2. Abstract

This thesis is an investigation into how to exploit the unique features of computer networks (notably the Internet) to support self-organising groups of adult learners.

The structures of systems influence the behaviour of their parts whilst those structures are in turn influenced by their parts’ interactions. Effects of structural hierarchies in popular systems of education may lead to poor learning experiences for some students. An alternative way of organising such systems is to decentralise control and to allow a structure to emerge from the combined actions of learners: a self-organised learning environment.

The functionality of a teacher often has the largest effect on the dynamics of an educational system. This is therefore a good place to concentrate efforts to encourage emergent structures to develop in adult education. The thesis attempts to classify of what that functionality consists, abstracting the roles a teacher may perform.

The Internet (especially the World Wide Web) has more of a network than a hierarchical structure and, being a virtual space, provides relatively virgin ground on which a less centralised model of educational organisation might develop. The thesis considers how self-organised learning may arise in existing Internet-based environments. It identifies a key weakness of existing systems to adequately address the varied and ever-changing needs of learners.

A number of studies performed as part of this investigation centre on the construction of a series of software products explicitly aimed at enabling the self-organisation of learners. They achieve this through the adaptation and evolution of metadata at different structural levels, thereby dynamically adapting to learners’ needs as those needs develop.

The thesis concludes with a set of guiding principles for those seeking to build self-organisation into learning environments.

 

Contents

3. Abstract *

   Contents *

   Table of figures *

   List of tables *

   Acknowledgements *

   Preface *

   Context *

   Research methodology *

   The process – a reader’s guide to the chapters *

   A foretaste of some conclusions *

   Chapter 1 : Organised education *

   About this chapter *

   Education *

   Beginnings *

   Characteristics of complex systems *

   Parts and wholes *

   Complex Adaptive Systems *

   A systems view of education *

   Individual learning as a complex adaptive system *

   Group processes *

   Control from the centre *

   Decentralised control *

   Distance learning systems *

   Independent learners and teams of teachers *

   Excessive complexity *

   Emergence of the unwanted *

   Conclusions to this chapter *

   Chapter 2 : How systems change *

   About this chapter *

   Evolution *

   Characteristics of evolving systems *

   A spandrel in the works *

   Parcellation drives evolution *

   Order for free *

   Stigmergy *

   Forests and trees *

   An educational hierarchy *

   Finding an appropriate sense of scale *

   Scale in educational systems *

   Conclusions to this chapter *

   Chapter 3 : Teacher features *

   About this chapter *

   Do we really need teachers? *

   Self-teaching *

   Peer tutoring *

   The perils of a chopsticks culture *

   So what do teachers do? *

   What is a theory of instruction? *

   Abstracting the roles of a teacher *

   A teacher is a subject expert *

   A teacher is concerned with communication and the mediation of communication *

   A theory of instruction should specify the experiences which implant a predisposition toward learning *

   A theory of instruction should specify ways a body of instruction should be structured for easy assimilation by the learner *

   A theory of instruction should specify the most effective sequences to present materials to be learned *

   A theory of instruction should specify the nature and pacing of rewards and punishments in the service of learning and teaching. *

   An holistic view *

   Other things a teacher might do *

   Conclusions to this chapter *

   Chapter 4 : Hillocks in the stuff swamp *

   About this chapter *

   Teachers are bad for learning *

   Finding good stuff *

   Brute force and ignorance *

   Discussion fora, flora and fauna *

   My own experiments *

   The first study: raw fora *

   The second study: Adding structure. *

   The third study: adding stigmergy *

   Conclusions to this chapter *

   Chapter 5 : Planting gardens in the stuff swamp *

   About this chapter *

   Structuring of content *

   Communication *

   Synchronous communication tools *

   Asynchronous communication tools *

   Computer Supported Collaborative Argumentation tools (CSCA) – e.g. D3E and POW! *

   Annotation systems – e.g. uTok, Comentor, CoNote *

   Knowledge bases – e.g. the Answer Garden, Knowledge Garden, Mecpol, Q&A sites *

   Recommender systems *

   Looking for the good stuff- descriptions of utility *

   Seals of approval *

   Collaborative filtering *

   Forms of collaborative filter *

   Some examples of collaborative filtering approaches to the discovery of resources and their value *

   Recommender systems explicitly designed for learners *

   Capturing knowledge *

   Problems with recommender systems *

   Descriptions of relationships *

   Pathfinding and trails *

   Assessment *

   Conclusions to this chapter *

   Chapter 6 : Burying Teachers in the Stuff Swamp *

   About this chapter *

   Problems of Categorisation *

   Information Retrieval issues *

   Novelty *

   The Motivation for CoFIND *

   CoFIND as it is now *

   CoFIND as a replacement for the teacher *

   A theory of instruction should specify the experiences which implant a predisposition toward learning *

   A theory of instruction should specify ways a body of instruction should be structured for easy assimilation by the learner *

   A theory of instruction should specify the most effective sequences to present materials to be learned *

   A theory of instruction should specify the nature and pacing of rewards and punishments in the service of learning and teaching. *

   Subject expertise *

   Community *

   Reflective practice *

   Conclusions to this chapter *

   Chapter 7 : CoFIND developments and findings *

   About this chapter *

   Generations of CoFIND *

   Methodology *

   CoFIND 1 *

   The problem *

   CoFIND functionality *

   CoFIND 1 case study *

   Some Results *

   Conclusions for the CoFIND 1 study *

   CoFIND 2 *

   Topics *

   Islands and bridges *

   Collaborative trails *

   The system in action *

   Simon’s Clients and Severs [sic] *

   Issues still to be resolved with the topic mechanism *

   Conclusions *

   Qualities *

   Four studies *

   Algorithms for qualities *

   WebSite Evaluation System *

   The problem *

   The process *

   Evaluation *

   Conclusions *

   N- vs. 1- dimensional evaluations *

   The problem *

   Design *

   The process *

   Conclusions *

   Stigmergy and qualities in CoFIND 2.6 *

   The problem *

   Design *

   Qualities red in tooth and claw *

   Evaluation *

   Conclusions *

   The effect of topics on qualities *

   The problem *

   Selecting appropriate topics to compare *

   Comparison of the use of qualities for different topics *

   Evaluation *

   Conclusions *

   Note: the posting and creation of online documents *

   Current limitations and future horizons *

   Was there any real point to these exercises? *

   Does it help people to learn? *

   The cold start *

   Skyhooks *

   The ageing of resources *

   Scheduling *

   Tasks and synchronisation *

   Shapes and forms *

   Clustering *

   Metadata are more than just words *

   Relationships in general *

   Conclusions to this chapter *

   Chapter 8 : Towards a theory of self-organising network-based learning systems *

   About this chapter *

   Where are we now? *

   Guidelines for the construction of self-organising network-based learning environments *

   Foundations *

   Growing self-organised systems *

   Creating structure through metadata *

   Helping learners to learn *

   A self-organised system *

   Independent and original contributions to knowledge *

   Conclusions *

   Appendix: related publications *

   Bibliography *

   Index *

Table of figures

Figure 1.1 – research methodology overview *

Figure 2.1 – detail of Canaletto’s Interior of San Marco showing spandrels *

Figure 2.2 – an educational hierarchy *

Figure 4.1 – messages with stigmergic ratings *

Figure 5.1 – JIME *

Figure 5.2 – a typical web-board *

Figure 5.3 – uTOK *

Figure 5.4 – information flocking *

Figure 6.1 – feedback loops within CoFIND *

Figure 7.1 – CoFIND 1 quality selection screen *

Figure 7.2 – CoFIND 1 resource display screen *

Figure 7.3 – CoFIND 2.6 topics (network management group) *

Figure 7.4 – CoFIND 2.6 topic selection screen showing self-organisation (perhaps) *

Figure 7.5 – CoFIND 2.6 topic selection screen, showing Simon’s clients and severs [sic] *

Figure 7.6 – CoFIND 2 resource display screen *

Figure 7.7 – stigmergy and qualities *

Figure 7.8 – CoFIND usage throughout the year (MSc group) *

Figure 7.9 – comparison of the use of qualities for different topics *

6. List of tables
7. Table 3.1 – Bruner and Gagné juxtaposed *

   Table 4.1 – thread structure in the MSc networking assignment newsgroup *

   Table 7.1 – the main iterations of CoFIND *

   Table 7.2 – use of qualities over time for the HCI module *

   Table 7.3 – the effect of the use of qualities on ratings for the network management group- overall *

   Table 7.4 – the effect of use of qualities on ratings for the network management group- by quality *

   Table 7.5 – the effect of the use of qualities on ratings for the MSc group- overall *

   Table 7.6 – the effect of use of qualities on ratings for the MSc group- by quality *

   Table 7.7 – the effect of pages on click-throughs *

   Table 7.8 – the lack of effect of ratings on click-throughs *

   Table 7.9 – influence of pages on click-throughs *

   Table 7.10 – ratings following a visit to a given page for the network management group *

   Table 7.11 – The effect of page number on click-throughs for the MSc group *

   Table 7.12 – the effects of stigmergy on quality selection *

   Table 7.13 – different uses of qualities for different topics *

10. Acknowledgements

2. Preface
1. Context

Although much of this thesis revolves around education, it will not be much concerned with theories of learning, but instead with the kind of environments in which learning happens. In Socratic terms, its objective is to discover what makes a good midwife, not a good baby. Although the two are intimately linked, I believe that it is possible to meaningfully consider either in relative isolation.

Mostly this thesis will be concerned with adult learners and my own experiments have been conducted entirely within the confines of a traditional university system. However, when discussing historical and theoretical issues the boundaries will blur from time to time with issues that affect other kinds of learner and different forms of learning.

2. Research methodology
1. The main research question

This thesis is the result of a search for principles and guidelines for the construction of network-based learning environments where organisation and control arises from the combined actions of learners using networks rather than from designers or managers of learning. The big question is therefore:

how should self-organised network-based learning environments best be constructed?

The answers I shall give to this question will be based on an investigation of the nature of educational systems, of self-organising systems in general and of instances of self-organising systems on the Internet. Above all, they are based on the ongoing development, use and evaluation of a piece of Web-based software known as CoFIND (Collaborative Filter in N Dimensions), which is designed to facilitate self-organisation of learning by providing an environment created by and adaptable to the learners who use it..

The main research approach for the bulk of the thesis is based on action research, as defined by Burns as:

the application of fact-finding to practical problem-solving in a social situation with a view to improving the quality of the action within it.

2. Why action research?

At the outset my research questions were fuzzy. I knew that I wished to explore ways that evolution could help groups of learners on the Internet manage their own learning, but beyond that I had little idea of the specific ways that this might occur. With its focus on clarifying and exploring, action research provided an appropriate means of helping to clear up this fuzziness.

Given the fuzziness of the questions I have used predominantly qualitative methods as my key objective is to generate hypotheses. Burns quotes Barton and Lazarfield as describing a qualitative approach thus: "like the nets of deep-sea explorers, qualitative studies may pull up unexpected and striking things to gaze on." Mainly through a process of building and evaluating computer systems I have been seeking the unexpected, which by definition will not succumb to a scientific method of inquiry and analysis. The central purpose of this research is discovery, with the forms of explanation supporting and determined by the results of discoveries as they arise:

Action research does not know what questions to ask until it has interpreted the present.

.

In the process of performing my research I wished to create usable tools and methodologies to assist my teaching. A key strength of action research is that it is generative. My intention has been to create as much as to describe or explain. I therefore needed a set of methodologies that was responsive, adaptable and creative. Cohen & Mannion characterise the key goal of action research as "to add to the practitioner’s functional knowledge of the phenomena she deals with."

I wished to combine my research with my teaching. Action research is an effective approach for improving and extending both. Braa and Vidgen observe that action research can be seen as "a way of building theories and descriptions within the context of practice itself". It is also a feature of this approach that I have been bound by the ethical demands of my profession as a teacher. I could not reasonably perform experiments on my students which I did not perceive to be in their own interests. This usually requires a greater degree of intervention and involvement, invalidating any claims to scientific truth.

Much of my research revolves around the creation and use of CoFIND. CoFIND is an exercise in model building, an attempt to make use of principles of self-organisation to achieve form and structure. Some people build models to demonstrate things they already know, like those perfectly constructed Lego houses based on the plans that come with the box. Others throw away the instructions and begin to build. As the Lego models develop, so their earlier features determined their later form, usually leading to unanticipated constructions which might be inspired or simply awful. This approach underlies the design of the CoFIND system. Through the process of building and testing I have explored possible forms of educational self-organisation by trying to encourage them to happen. I have been testing hypotheses through the construction of systems but at the same time using those systems to generate hypotheses.

Each CoFIND system has incorporated the ability to measure many aspects of how it is being used. This ability to self-document is a notable advantage of working in virtual space and allows examination of the use of the system in ways that would be difficult in a traditional face-to-face environment.

I have made use of techniques which seem appropriate for each stage of my research. For example, in investigating the interactions in a newsgroup I adopt a broadly ethnographic approach, while in analysing the effects of some aspects of my system I employ a quasi-scientific method, comparing control groups to identify differences in their behaviour.

3. Details of my approach

1. What I have not done
1. An action research variant

Although I have described my approach as based on action research, I have deviated from standard practice in a number of important respects. In particular, most of what has been achieved is not a team effort, although it is concerned with changing the nature of communication between learners by adding to their repertoire of possible ways of interacting. I have used it more as a development methodology for a computer system, with theory spinning off as a significant side-effect.

2. Quantitative research

In taking on action research as an approach I have made a number of trade-offs, most notably in generally avoiding quantitative methods, thereby perhaps reducing the generalisability of my findings. The lack of scientific validity seems a small loss compared with the generative gains and the learning which is entailed.

3. Measurement of effects on learning

1. The process – a reader’s guide to the chapters

In chapter one, based largely on existing literature, I argue that most educational systems result in and from complex systems behaviour. This behaviour affects the constituents of educational systems in ways that are sometimes antagonistic to learning. Few of these observations are original, though because of the purpose behind them (to provide a theoretical foundation for building self-organising learning environments) the specific way in which they are combined is unusual.

Chapter two investigates ways in which complex systems arise in general, drawing on theories surrounding evolution and complexity to lead to a discussion of how such features might relate to existing educational systems. I arrive at conclusions about the features of such systems which should be examined in order to produce self-organising versions of them. In particular, I conclude that the most interesting results are likely to be achieved if we concentrate on the role of the teacher within any system that we build.

Chapter three follows from the conclusions of chapter two and is intended to abstract significant features which characterise the roles of the teacher or manager of learning. The purpose of this is to identify what it is that we are aiming for in a learning environment.

Chapter four looks at systems which may have (or may be adapted to) some educational purpose and which are not designed with any inherent self-organising characteristics but which none-the-less, because of the ways that individuals use them, exhibit self-organised behaviour in greater or lesser degrees. The chapter includes a piece of research which attempts to identify ways in which self organisation is possible within a system not explicitly designed to organise itself. This takes the form of three related studies in which I applied a series of interventions to identify some of the effects of structure and stigmergy (communication via the environment) on the dynamics of a discussion forum.

Chapter five looks at Internet-based systems with features which are either explicitly designed to enable self-organising behaviour or which exhibit that behaviour because of implicit mechanisms built in to them.

Chapters six and seven describe the reasoning behind and the development of a group of systems (CoFIND, Collaborative Filter in N Dimensions) which are explicitly designed to enable self-organisation to occur. Chapter six explains the operation of and the rationale behind construction of the system as it currently stands. Chapter seven is a detailed explanation of the interventions I made and the effects that these had on ongoing development. CoFIND was used in as many different situations as possible, in the hope that it would be possible to correlate and compare results. Although this was not a process of triangulation, the existence of similar systems in different contexts helps to improve confidence in the accuracy of the results.

The final chapter summarises the lessons learnt and attempts to extract a set of rules for building self-organising learning environments from the results of the foregoing chapters.

2. A foretaste of some conclusions

2. : Organised education
1. About this chapter

In this chapter I will be looking at educational systems and the importance of structure in determining their nature. In general (as throughout this thesis) the emphasis will be on adult learners with a strong bias towards traditional academic systems typified by universities. However, where appropriate, I will occasionally extend the boundaries to provide a more inclusive view of education.

A part of the chapter will be devoted to attempting to understand how structure in systems develops in general. This will include an investigation of how the components of any system are affected by their relationships to each other and to negasystems, those features which fall outside the system under observation yet which interact with it. With this systems-oriented perspective in mind, I will look at educational systems at a number of scales, building up from the systemic nature of individual learning, to the dynamics of group learning, to the effects of complete educational systems on their participants. I will attempt to show that in many ways these effects can be positively detrimental to the desired outcome of learning. However, the field of distance education (especially when conducted using computer networks) offers opportunities to deconstruct then rebuild educational systems in different and perhaps better ways.

2. Education

1. Beginnings

1. Characteristics of complex systems
1. Parts and wholes

And I know not if, save in this, such gift be allowed to man

That out of three sounds he frame, not a fourth sound, but a star.

From Abt Vogler, Robert Browning

In the latter half of the twentieth century but with roots stretching back to the works of Charles Darwin and beyond, there has been a paradigm shift in scientific thinking. The success of traditional scientific ways of thinking was underpinned by a process of reductive abstraction and decomposition into primitive parts, from which all manner of theories and engineering marvels could be assembled. The power of simple laws and discoveries of the sorts identified by Newton, Galileo, Keppler, Voltaire, Rutherford, Einstein and a host of other luminaries has led to results as wide-ranging as space travel, mass transport, genetic engineering, disease control, atomic power, plastics, electrical power, and television. Reductive science works in a lot of different contexts. However, it has gradually been realised that (perhaps) the majority of interesting phenomena are not susceptible to reduction of this sort. Much of what affects us most deeply is the result of the behaviour of systems, large-scale behaviours caused by small-scale interactions and (just as interestingly) complex interactions leading to simple behaviours. Cohen and Stewart call these, respectively, simplexity and complicity .

Systems thinking covers a broad range of related schools of thought, concepts and theories, including evolution, cybernetics, chaos theory, system dynamics, fuzzy logic and the science of complexity and simplicity, collectively christened by Murray Gell-Mann as ‘Plectics’ . The underlying strand of thought that relates these various fields is the idea that the behaviour of all systems follows common principles. We can see "patterns that connect" in systems as diverse as computers, animal bodies and cells, population growth, financial systems, the spread of ideas, the growth of cities, central heating systems and cloud formations. We even see them in systems of education.

We are going to be looking at systems without governors or central controllers, which have properties of emergent order. That is to say, we are looking at systems which are more than the sum of their parts, as the chord in Browning’s poem is more than the sum of its three constituent sounds, a point taken up by the Gestalt theorist Wertheimer , who writes on the subject of how notes form a melody:

What I really have, what I hear of each individual note, what I experience at each place in the melody is apart which is itself determined by the character of the whole. What is given me by the melody does not arise (through the agency of any auxiliary factor) as a secondary process from the sum of the pieces as such. Instead, what takes place in each single part already depends upon what the whole is. The flesh and blood of a tone depends from the start upon its role in the melody: a b as leading tone to c is something radically different from the b as tonic. It belongs to the flesh and blood of the things given in experience [Gegebenheiten], how, in what role, in what function they are in their whole.

The notion of the relationships between the parts forming a whole which is greater than the simple sum of those parts is central to a view of the world from a systems perspective . Systems are recognised and defined by their connections, patterns, interdependencies and emergent behaviour. By identifying repeating patterns within different systems we may see connections between different subjects and disciplines which transcend their subject matter.

2. Complex Adaptive Systems

There are widespread classes of systems which have been identified by Holland as Complex Adaptive Systems (CAS). Prigogine described these as dissipative structures, open systems which exchange matter, energy or information with their environment and which thereby exhibit order . Examples might include tornadoes, whirlpools, population changes, movements of money markets or evolution. Prigogine demonstrated that only closed systems move towards entropy, whereas open systems move towards some form of organisation. A CAS exhibits a range of interesting properties, particularly that it generates emergent properties of structure. Such properties are not predictable from any individual input but are the result of the behaviour of the system as a whole. Self-organising systems are those where patterns form and acquire order and structure through interactions internal to the system . They do so through a combination of positive and negative feedback mechanisms, the positive feedback providing change, the negative feedback stopping that change occurring. The patterns which arise do so not because of a high level rule but as a result of individual actions within the system. Camazine et al give the example of the formation of colonies of birds, all of which follow the rule "I nest where you nest" (positive feedback) "unless it is too crowded" (negative feedback). The result is separate colonies. This is an example of what is meant by emergent behaviour. There is no grand plan which the birds are following to form a colony. It just happens because of positive and negative feedback rules, which determine the structure of the system in terms of the relationships between its parts.

2. A systems view of education

Considering educational systems from a systems perspective can yield important insights. As Senge has persuasively argued, structure influences behaviour. A wide variety of behaviours are born out of the nature of systems, not the individual constituents of those systems. Senge gives an extended example in the form of the Beer Game, where retailer, wholesaler and manufacturer are caught up in a vicious cycle of supply and demand which, when not viewed in a systemic way, leads inevitably to cycles of over-ordering, depleted inventories then heavy over-stocking. The problem arises out of the nature of the system, which includes varying degrees of latency between deliveries. Retailers, responding to a consistent slightly increased demand for a certain brand of beer, increase their orders from the wholesalers. The wholesalers increase orders from the brewer. Because of inherent latencies within the system, partly due to delivery cycles and partly due to how long it takes to brew beer, orders are not immediately met, so everyone perceives increased demand and an inability to meet it. To compensate, the retailers increase orders to cope with the backlog, the wholesalers follow suit a little while later, the brewer brews more beer to cope. These effects are magnified as they rise through the system. Eventually, the backlog of orders are met, at which point all players in the system have a massive over-stocking of beer, orders are cancelled and all players suffer accordingly. This game demonstrates system generated behaviour of the sort which can be seen in the wildly fluctuating availability and pricing of microchips, for example. Senge gives many more examples of such unwanted artefacts of system behaviour, including arms races, bank runs and fuel crises, where individual decisions are made based on rational grounds as a result of a perceived stimulus yet the overall effect of the system acts to the benefit of nobody. The nature of such systems is to cause emergent behaviour which is not predictable by examining a part or parts of the system in isolation. Positive feedback loops (which create) and negative feedback loops (which constrain) drive systems to a state of self-organisation. This self-organisation can occur spontaneously even in simple systems. Chaos theory shows us many examples where strange attractors emerge out of apparent chaos, like the rhythms of dripping taps, the patterns of snowflakes, or Jupiter’s red spot,

a self-organizing system, created and regulated by the same non-linear twists that create the unpredictable turmoil around it.

A good example of the importance of structure in determining the fate of its parts lies in the concept of an ecosystem. Here, a delicate and always shifting balance of predators, prey, consumers, producers and the environment in which they live (and often, as notably displayed in the case of rainforests for instance, create) determines the forms of all the organisms within it. No single factor can account for the forms which arise, but the overall structure results in distinctive flora, fauna and all the other myriad taxa of life within a given area.

As this thesis progresses, I will attempt to identify ways that we can make use of such structures to achieve positive educational benefits, but for now I will look at the ways that existing educational systems are constructed.

1. Individual learning as a complex adaptive system

Cafolla and Kaffman provide an elegant proof that the very process of learning is non-linear. They simplify the learning model by assuming that the state of new information received is related to the state of information already learnt, and that there is a certain (unspecified) amount of degradation that occurs in the information retained. The combination of these factors leads to a simple equation (x+1=L xi(1-xi)) which, when graphed, shows chaotic patterns of sensitivity to initial conditions. Thus, even without the massively large number of potential influences on learning, as a self-contained occurrence it may be seen as a complex adaptive system.

Piaget has exerted a strong influence on educational thinking for many years as one of the main debunkers of the behaviourist paradigm of education which still exerts its unholy influence on the processes of formal education. It is interesting that his structuralist perspective coincides somewhat with the ideas of systems thinking. Piaget writes:

As a first approximation, we may say that a structure is a system of transformations. Inasmuch as it is a system and not a mere collection of elements and their properties, these transformations involve laws: the structure is preserved or enriched by the interplay of its transformation laws, which never yield results external to the system nor employ elements that are external to it. In short, the notion of structure is comprised of three key ideas: the idea of wholeness, the idea of transformation, and the idea of self-regulation

Hence, the underpinnings of Piaget’s ideas are firmly rooted in the notion of learning being a complex and dynamic system with feedback at its heart. Piaget’s form of structuralism has much in common with the ideas of Goodwin and other complexity theorists who identify the emergence of form as a separate (if complementary) issue to that of evolution. For Piaget, structure is an emergent property which feeds on itself and its environment to maintain its stability through assimilation, adapting and changing into a new structure by a process of accommodation. In other words, a disturbance to this system will result in the transformation of old structures into new structures. For Piaget, it is this process that leads to learning. If we are to consider this in relationship to the structure of teaching, we can see that in order to achieve learning we must provide resources which provide knowledge which can be assimilated and, as with Vygotsky and his zones of proximal development, aimed at just over the edge of that boundary between the continuation of the old structures and the creation of the new.

2. Group processes

Piaget’s and Vygotsky’s models of learning stress continuous interaction with the world and other people. Gordon Pask’s conversation theory places this process of communication at the heart of learning. He is also an early proponent of viewing educational systems in their entirety, not divorcing teaching or learning or administration or any other factor from his systems approach. Pask’s view of the systemic nature of education considers the individual as a subsystem within a system. Rocha explains Pask’s view of the internal system of learning and its relationship to the educational environment thus:

[Pask] developed an extensive theory of conversation that proposed the abandonment of the concept of learning as a one to one mapping of real world to mental categories, for a dynamic, internal, self-organizing process of coming to know, constrained by developmental interaction with an environment and fellow ‘knowers.’

For Pask, learning is truly situated and contextual, an iterative process of construction which incorporates previous knowing with new knowledge in a process of interaction between the learner and other learners, the learner and the teacher and the learner and the learning resource.

In "Teaching Strategies: a systems approach," Pask & Lewis analyse four teaching approaches in terms of systems thinking, looking at the extent to which they make sense as an educational system. In systems terms, these can be divided into those which exhibit central control and those where control arises from interactions of individuals within the system. The characteristics of systems which are controlled from the centre have a basis in the reductive theories of the behaviourists, whereas those which are more distributed borrow more from systems theory and lead inexorably to a variation on constructivism.

3. Control from the centre

Pask & Lewis describe two centrally controlled systems for teaching, the behaviour shaping and the reinforcement menu curricula. In these approaches to teaching, "the learner does things and the things he does are influenced by things done to him" . This method of design deals in measurable behavioural outcomes. It is much the same method that Houghton equates with "an engineering approach of scientific management to the problems of education," not in keeping with the complex, situated nature of education. A central problem with this approach is that the behaviourist cannot say with certainty that the learner has a concept, merely that the learner acts as though he or she has a concept. The measurement of this behaviour is problematic and does not necessarily lend itself to generalisation. To take a trivial example, the ability to divide 100 by 10 does not necessarily transfer to the ability to divide 101 by 11. Motivation is provided by rewarding the learner (rather like Pavlov’s dogs) to reinforce learning goals. In this model the teacher is a simple regulator, taking a list of desired facts or behaviours and rewarding the student (usually with good grades or words of encouragement) for displaying those behaviours. This model, though common, does not often result in deep knowing:

It is not unusual to find learners, with perfect terminal behaviours, who cannot answer a single generalization question. Nor is it very unusual to find that a learner who has successfully completed a programmed text fails dismally if interrogated the next day.

This view is echoed by Houghton, who claims that further consequences are that such an approach "renders the unplanned unlikely....eliminates innovation ...is crashingly dull, uninspiring, and unmemorable except for its boredom." . A similar point is made by Saba , adding the important insight that, even where such an approach is successful, it is inappropriate to student needs:

this model of instruction might have been useful for uniform industrial processes, when standardization of tools and manufactured products was a challenge and a desirable goal to achieve. In the foreseeable future, the challenge for most front-line workers will be how to deal with unpredictable situations with novel solutions.

The application of reductive scientific methods leads to an industrial paradigm of measurable inputs and outputs. Ramsden writes:

the significance of independence and choice emerges repeatedly in research on student ratings and perceptions of favourable academic environments, at higher and upper secondary education levels. Yet most prevailing systems of learning in higher education adopt mass-production standards; they handle each individual student in the same way, even though we know for certain that they operate in different ways.

A proud flower of this reductionist paradigm is the external examination, which Whitehead rightly defames as ‘fatal to education’ . The use of summative assessment lies at the heart of much of formal education and creates some of its greatest weaknesses. By attempting to stamp a particular body of uniform knowledge onto the learner and to set rigid criteria by which the learner can be measured, the system makes motivation extrinsic to the task and so disassociates it from the desire to learn and explore. Holt puts it well:

when we reward children for doing what they like to do – find out about the world- they soon learn to do it only for rewards. Since the rewards of school only go to a few winners, most children, the losers, stop asking questions. This is one of the flaws in the idea of positive reinforcement; it only works as long as we keep it up.

The examination becomes the end, not the process of learning. The point was equally well made over one hundred years ago by Dewey, who wrote:

I believe that much of present education fails because it neglects this fundamental principle of the school as a form of community life. It conceives the school as a place where certain information is to be given, where certain lessons are to be learned, or where certain habits are to be formed. The value of these is conceived as lying largely in the remote future; the child must do these things for the sake of something else he is to do; they are mere preparation. As a result they do not become a part of the life experience of the child and so are not truly educative.

It is disappointing that this point still needs to be made. Like Pask, Dewey’s view of the educational process is one primarily rooted in communication and, like Pask, Dewey sees the process as extending beyond the barriers of a simple fact-laden curriculum.

4. Decentralised control

As alternatives to centralised control, Pask & Lewis describe the processes of guided discovery and conversational learning curricula, both part of a constructivist ethos aimed at the construction of an elaborate network of concepts, or structured knowledge. The learner is seen as a problem solver, with a rich level of interaction with colleagues and teachers towards that goal. In the guided discovery curriculum the teacher exerts more control, allowing the learners to construct their own learning but within the constraints and towards the end intended by the teacher. In the conversational learning curriculum (Pask’s paradigm of good practice) the learner may potentially take many paths to reach the goals, allowing knowledge to be structured differently from one learner to the next. The conversational process encourages reflection on the learning process, with the teacher acting only as a kind of help desk, albeit an active one concerned with asking questions and pointing out possible directions. As a result, the learner is able to form links between concepts, ideas and practice and is able thus to accommodate knowledge into existing schemes more effectively and to apply knowledge in a variety of contexts. A complex set of feedback loops results in a system where both teacher and learner adapt. In both guided discovery conversational learning curricula the interactions between learners, subject matter and teacher are significantly more complex than behaviour shaping and reinforcement menu curricula, with a rich variety of interactions (questions, answers, guidance, confirmations) at different points in the iterative learning process. There is a positive feedback loop between teacher and student in this conversational process, so that the teacher has the power to adapt teaching to the learner’s needs. Pask & Lewis write "the nature of a curriculum scheme clearly hinges on the sort of system a learner is supposed to be" . Through this style of curriculum the learner comes back into the picture as an independent and actively influential entity. Instead of a closed feedback loop of incentive and reward acting upon the student (the teacher merely acting as a kind of thermostatic control) the main agents in the system interact with and affect each other in a manner resembling an op-amp circuit, driven by positive feedback. As well as negative feedback loops there are also positive loops, which lead to creativity, motivation and emergent knowledge.

In common with many educational theorists Pask is abstracting idealised behaviours from what are always (as his own analysis clearly shows) more complex systems than can be characterised this way. In real life, it is very hard to find a pure behavioural model of teaching with no feedback between teacher and student, even if this is only in the form of a yawn from the back of a lecture theatre, As Wiener tells us:

there is no task harder for a lecturer than to speak to a dead-pan audience. The purpose of applause in the theater- and it is essential- is to establish in the performer’s mind some modicum of two-way communication.

There are also feedback loops extrinsic to the subsystem of teacher and learner, such as quality assessment bodies, the work of educational theorists and the interactions between students outside the classroom, which make even the dullest of behaviourist curricula part of a more complex set of feedback loops. It is equally rare to find a pure model of constructivist teaching which involves no centrally generated prescription of form and content, nor extrinsic incentive or reward. However, Pask’s way of seeing the world in terms of interacting agents gives a strong indication of what features a complex learning system needs. In the guided discovery and conversational learning curricula, there is a move away from central control towards a feedback model where all participants in the system are active agents. As a theoretical basis for a self-organising learning environment, therefore, the conversational framework that Pask proposes seems powerful and persuasive.

There is a close correspondence between Pask’s concepts of guided-discovery and conversational models, and the structured and emergent collaborative events discussed by Nachmias et al . Structured collaborative tasks are "well planned didactic solutions that follow the conceptual guidelines proposed by collaborative learning researchers and practitioners." They are therefore subject to central control. Emergent collaboration, on the other hand, is "the process by which group configurations and transactional patterns evolve among participants during the course of learning in a specific course." Tasks are defined in terms of goals and constraints, with the implementation of solutions involving potential shifts in perspective, fluid sub-groupings and varying roles. This perspective is helpful in providing a structure in which to create innovative and student-led environments where (for instance) students collaborate to create a knowledge base of educational resources according to agreed criteria, or suggest uses of discussion groups based on their experience. This sort of direct iterative feedback mechanism should be seen as a vital part of the process of constructing self-organising learning environments. Nachmias et al report on student moderated discussion groups, noting that student moderators "invested a lot of time and creativity to create a contextual framework to the discussion, and used email to encourage their classmates to participate." The fact that students were able to effectively perform this role is encouraging. It suggests that one of the more ‘human’ of teacher roles might well be performed by untrained learners. However, the report notes that teachers were still available to guide and support the moderation groups’ work, which implies that the roles were not emergent but structured from above.

Pace Marshall draws lessons from the paradigm shift from Newtonian science to systems thinking and complexity. She observes that we have constructed our educational systems "by drawing lines and boxes around everything and by separating things into discrete observable, measurable categories. We created dichotomies, divisions, departments, boundaries, and closed systems." . This has resulted in what she describes as "learning-disabled institutions" which have suppressed creativity and challenge "the personal, active, volitional, and social dimensions of learning that are so essential to authentic meaning." Her rally call is that:

As complex learning systems, schools are far more organic and dynamic than linear. We, therefore, must design them to function less like clocks, and more like kaleidoscopes.

This organic view of education resonates with a number of the themes that characterise learning theories of the twentieth century, where learner-centredness implies that control is derived from the periphery, not the centre. This is in stark contrast to the industrial model which still sits at the heart of all major educational institutions from schools (with their national curricula, planned lessons, timetables and specific targets) to universities which still insist on the bizarre practice of expecting hundreds of students to learn the same thing in the same place at the same time with little or no active involvement or communication, even though there is overwhelming evidence that learning occurs despite rather than because of these actions. There is even a case to be made for suggesting that what actually forms learning is what happens at the periphery of these events, the (sometimes informal) surrounding activities of communication and active experimentation and construction. If we are to take the constructivist ethos to heart, it is only in actively making connections and building knowledge on an edifice of previous knowledge that deep learning takes place.

Robert S. Houghton’s PhD thesis argues that an emergent paradigm (based on principles that allow for self-organisation, sensitivity to initial conditions etc) better characterises the educational process than the industrial model:

Perhaps the principle conclusion derived from chaos theory and the larger emergent paradigm for education is simple and basic: interact. To interact is a fundamental directive for education as a system and the individual as a learner.

In keeping with our systems view of education, for Houghton interactions form the structure of learning. Any system is defined by the interactions of its elements: these interactions make it a system in the first place, as opposed to a group of static elements. Dewey believed that this was the main thing that needs to be said about education . Hunter and Benson similarly argue that "all learning and all knowledge is the result of different realities co-existing in the world, at one time and that we create these realities through social interactions " . However, the degree of interaction and the types of interaction are still very much up for discussion.

David Pines, a researcher at the Santa Fe Institute, provides some fascinating insights into the ways in which universities exhibit the characteristics of complex systems at many levels. He identifies that "universities may be viewed as a collection of interacting communities each of which functions as a complex adaptive system" . It is instructive to note the many levels of self-organisation that Pines identifies. An educational system is not just a single entity but a collection of systems which themselves are collections of systems. He cites interactions between disciplines within faculties and between faculty and students as clear examples, although notes that the latter interaction has an unfortunate tendency to be one-way.

Different interactions result in different emergent structures, with varying degrees of order and chaos. The IFETS (International Forum of Educational Technology & Society) electronic discussion from 1-10 November 1999 was centred around the importance of Moore’s transactional distance theory. In the pre-discussion paper Lowe pointed out that there was a danger that too great a move to dialogue might lead to the original learning objectives being compromised . Communication is a necessary but not sufficient condition for achieving learning through teaching. Although it can provide positive feedback loops which generate new knowledge and connections, it can run out of control or dissipate. Saba noted that "instructional systems design must include negative feedback loops to optimize interaction at a distance and keep learning sessions productive" . It is not enough for communication to take place, but at least some of that communication has to be affective. Systems are constituted in the relationships between their elements. If elements to not affect each other, we are not looking at a system but at a set of discrete parts.

Like Pask, Diana Laurillard views the educational process as one involving an interacting loop of conversations between its key participants. These interactions constitute a set of feedback loops which implicitly leads to a self-organising system, involving teachers, students and learning resources. It is informative to distinguish this from such theoretical systems as the Kolb cycle, which characterises the learning process as a cycle of abstraction, reflection, experimentation and concrete experience. This cycle says nothing about the control mechanisms which operate to produce it. Laurillard’s conversational framework is explicit about the effects that each actor in the system has on each other.

given ... the integrative nature of the learning process, the inseparability of knowledge and action, and of process and outcome, there is no logical ordering of parts of the process, as each part is constituted in its relation to the other parts

It is impossible to consider isolated features of an educational system- changing one part changes them all. This is a feature which may affect virtually all systems, educational or otherwise. By its nature, any system is constituted from the relationships of its parts. Changing parts or their relationships alters the behaviour of their nearest neighbours. If those neighbours are altered then they in turn affect their neighbours and so on throughout the system. Sometimes these effects will be small, sometimes they will just flip the behaviour of the system to another similar state, sometimes they will change the system utterly.

3. Distance learning systems

Many authors have recognised that it is not possible to consider a part of an educational system outside the context of subsystems which it contains and other systems with which it interacts. Of these, Michael Moore plays a key role in the growing perception of education as a total system. In particular he is interested in distance learning, especially significant due to the focus on network-based learning that informs this thesis. Distance learning is different. Moore is keen to emphasise that distance learning represents a departure from the ‘craft’ approach to teaching:

A distance education system should be thought of as a network of knowledge sources, processors, managers, communication media, and learners

The interdependency of each part of the system means that making a change to one part means making a change to the whole:

For example, the exact nature of the design, the communications technology used for delivery, and the interaction depend on the sources of knowledge, the student needs, and the learning environment in the particular course. Selection of a particular delivery technology or combination of technologies should be determined by the content to be taught, who is to be taught and where the learning will take place. Design of the instructional media depends on the content, the delivery technology, the kind of interaction desired, and the learning environment. All these will be influenced by policy and management. Furthermore, changes in one component of a distance education system have immediate effects on all of the other components

A complex set of interactions is typical of a system with emergent properties. Moore’s notion of transactional distance is similarly systems-oriented. The essence of this theory is that there is a linear relationship between the amount of communication (dialogue) involved in a distance learning course and the amount of structure and associated detail. The influence of this theory has been quite widely felt, presenting one of the first coherent theories of distance education as distinct from co-located education.

Saba has refined Moore’s theory of transactional distance and tested it experimentally, using a subtle form of discourse analysis to demonstrate the interactions of structure and communication . Saba is one of the few authors to make the link between this explicitly system-oriented view of distance learning and self-organisation:

The concept of self-organization is central to the understanding of distance education, which is a radically different kind of educational organization than the place- and time-bound systems currently at work in most educational institutions.

Distance learning, by changing the nature of the surrounding systems, changes the nature of the processes of education. It Saba is right, it may hold the key to the central theme of this thesis and so is an area worthy of examination in greater depth. Distance learning encourages a re-examination of the roles and functions that enable learning to occur.

1. Independent learners and teams of teachers

Traditional co-located courses and methods of training have tended to extract learners from their everyday contexts, fill them with knowledge and send them back out again, like cars being serviced. Although this model is still prevalent, it has been joined by a more flexible approach to teaching and training which dates back to the middle of the nineteenth century, when Isaac Pitman created the earliest correspondence-based distance learning courses. A key feature of these courses has always been that learners could choose when and where their learning takes place . Institutions provide structured course materials and asynchronous feedback, allowing the students to exercise a great deal of control over the learning process. Pitman’s system and those it inspired came to be known as ‘independent study’, a clear indication of the importance of the learners’ control over their own learning. This close association of distance learning and independent study pervades the whole area, although the association is not essential. For example, a learner given tuition by ‘phone or video-conference could hardly be said to be independent and indeed there is a sense in which a distance learning course, through its structure and design, provides very little more independence than a traditionally taught course. This is where Moore’s theory of transactional distance comes in: it is not the physical distance which is significant but the amount of communication. Where there is more structure there will be less communication (or dialogue) and vice versa.

The realm of distance education has had a profound impact on the ways in which we approach the business of learning. Although some functions of the traditional educational system are farmed out to others (procurement and management of buildings, books, food etc) and establishment of curricula and their associated examinations are often centralised, the delivery and much of the form of the education experience is largely left up to the individual teacher. Distance education provides an opportunity to rethink that model, with the necessity of deconstructing it to identify what is required to achieve effective learning. As a result, it has long been a fruitful area of innovation and careful analysis of roles, needs and requirements.

The notion of independent learning firmly took hold in the 1960s and 70s with the establishment of a variety of open universities, the most famous of which was and is the Open University (OU). The OU owes a significant debt to Charles Wedemeyer, director of the Articulated Instructional Media project whose early experiments and theoretical underpinnings played a large role in its inception and form. However, the OU has been hugely influential in its approach to the educational process. Moore & Kearsley suggest that what distinguished it was a "total systems approach to distance education" . The OU is fairly big by most standards applied to educational institutions, with around 130,000 enrolling students per annum. Dedicated from its inception to distance education with no campus-based courses, Walter Perry and the other creators of the OU were able to examine the teaching process closely and abstract its functions, creating production teams more akin to theatrical troops or film production teams than the traditional hierarchies of individuals and self-contained groups characteristic of campus-based universities. By dividing up tasks the OU is able to maximise quality and to control costs due its size and consequent economies of scale. Moore & Kearsley provide a useful way of understanding the benefits and potential barriers to such a process, dividing the commonplace ‘author/editor’ model (where an editor or editorial team takes the work of a content expert and turns it into a course) from the ‘course team’ model, pioneered by the OU:

Each course is designed by a team which might consist of as many as 20 or more people, each of whom is a specialist. At the design stage of the ISD [Instructional System Design] process a group of academics, content specialists in a subject area, writes outlines of what should be taught in their particular specialities and engages in extensive discussions regarding the objectives and content of each unit and module of the course. These outlines and objectives are debated not only by the academics and others but by the whole team, including producers, editors and external consultants. Eventually, after three separate drafts have been presented, a study guide on that subject emerges.

As well as taking responsibility for the study guide in their speciality areas, the academics assemble books of readings, make audio- and videotapes, and design tests and supplemental materials, all with the assistance and guidance of specialists in these tasks. These technical experts include editors, graphics designers, radio and television producers, instructional designers, and librarians, even a specialist photolibrarian. A senior academic chairs the team and an administrator is responsible for ensuring that each task in the development schedule (which often lasts one or two years) is completed on time.

By splitting up the roles which are more often shoe-horned into an individual or arbitrarily spread across a group of individuals, it is possible to explicitly recognise a range of roles within the educational system. The roles divided up amongst the team at the OU are often played by a single individual in a more traditional environment, or a small group of simply differentiated or undifferentiated individual roles within a team. This may have positive as well as negative effects. On the positive side such a method will typically produce results more quickly than (potentially) the two years it can take to produce a complete OU course. On the other hand, it demands a great many roles of an individual, who may not possess the qualities required in equal measure for each role.

4. Excessive complexity

1. Emergence of the unwanted

Educational systems may be viewed as complex adaptive systems . We have already seen how they are composed of interlocking systems at a wide range of scales which are linked through positive and negative feedback loops. They are open systems, constantly fed with information and interactions. Like all complex adaptive systems, educational systems exhibit emergent properties, not always predictable by looking at the individual parts of the system. In the scenarios which we just looked at, our God's-eye view made the sequence of events possess a curious inevitability, but a close-up observation of any single point in the proceedings would not have told us this. Perhaps we could say in each case that those who might have made the changes could, with more information, have foreseen and prevented the bad consequences from occurring. Perhaps it is just a question of adequate controls and feedback loops. A more likely explanation is that individuals made perfectly rational decisions in relation to the circumstances in which they found themselves. A teacher faced with an injured child took the child to receive treatment. Parents seeing a school with a poor record did not want their children to suffer so kept their children away. Like birds in a flock, they applied simple rules, which led to a overall system behaviours which might have been recognised but would probably not have been avoided, unless all actors in the system were systems theorists.

Onar Åm analyses the survival of the school system in systemic terms, noting that despite its manifest failings, turning children with an "intense, almost obsessive urge to learn" into those who are apathetic and reluctant to learn, it continues to thrive. Åm sees this as a self-reinforcing process. By systematically repressing curiosity and enforcing learning through the erosion of values, beliefs and attitudes, students become apathetic towards learning, preferring entertainment before academia. Thus, students, teachers and parents come to see learning as something that must be enforced to overcome apathy, and so the cycle repeats. Åm claims "the survival strategy of the enforcement system is to produce massive empirical evidence of student apathy which suggests that enforced learning is a necessity." It does not have to be this way and Åm’s indictment ignores the many positive efforts made to encourage the joy of learning. However, his analysis of the school system is almost evolutionary, in a Darwinian sense. There is a cycle of reproduction, with variation and competition implied (Åm’s own paper is an example of this) and the evidence he gives of a self-reinforcing feedback loop which works to the detriment of the learners shows that evolution does not of itself result in a system best fitted to learning. It survives because it is good at surviving, not because it assists learning.

A similar point is made by Forrester:

In complex systems, there are many interconnecting feedback loops. A new policy, which is intended to solve a problem, causes reactions in other parts of the system that counteract the new policy. In education, that reaction may come from administrators, from school boards, from parents who do not want new experimental ideas tried on their children, or from budget pressures.

For those of us in academia, adult education generally equates to a significant part of what goes on in universities and colleges. These embody a group of systems that have evolved slowly (very slowly) over hundreds of years from early, largely monastic, beginnings. Centuries (even decades) ago, the university represented an efficient way of bringing learners into contact with like minded learners, with access to scarce resources and the learned who could in principle help them to make some sense of it all. There was simply no other way of bringing the knowledge and understanding of academics together. Similarly, it made sense to put expensive books and manuscripts in one place, where they could be accessible to those seeking knowledge. As Hutchinson puts it:

…it was the purely physical restrictions on access to scholarly authority (whether medieval monk or Oxford professor) and to the written and printed word that necessitated the creation of bricks-and-mortar centres of learning.

Alternatively, it might have simply been greed, like the Sophists of whom Plato coined the phrase "a paid hunter of the young" who are credited with the first recorded instance of mass instruction . These pragmatically generated institutions thus formed the basis of an ecology which has developed in ways that might have been predictable but which had little to do with the overarching requirement, to produce research and learning. The system sustains itself in many ways, not all of them healthy. Illich likens the self-perpetuating cycle which encourages teachers into their chosen profession to the myth of Oedipus:

Oedipus the Teacher, who "makes" his mother in order to engender children with her. The man addicted to being taught seeks his security in compulsive teaching. The woman who experiences her knowledge as the result of a process wants to reproduce it in others.

The following quasi-historical explanation of the growth of our present-day systems is almost certainly not completely true and is at best only a small part of the story. However, if there is the slightest glimmer of truth in it then the point is made that educational systems’ behaviour might lead to undesirable consequences. Consider this then as more of a parable than a stab at historical truth.

University systems developed around the simple requirement to provide geographically collocated learning. Degrees were invented as a means of rubber-stamping the learner as the learned and this required assessment. Since the institution had brought large numbers of students together, such assessments needed to be cost effective and easily and consistently applied. Thus evolved a range of examinations and essays which soon became more than a rubber stamp but instead were the central focus around which education revolved, forming a focus which often limited learning by providing second-order goals which actively detracted from the wishes and predilections of the learners themselves. Individual and small-group tuition, though recognised as desirable, became economically unviable, so the Lecture was born. Increasingly, through a set of interacting processes from within and without the institutions the educational system that we know and love today was born. Combined with other factors like the mechanistic world-view that inspired (and was inspired by) the Industrial Revolution, some behavioural psychology and improvements in mass-production techniques for books and journals, a system developed whose crowning glory was a method of teaching which, Ramsden observes, "is actually detrimental to the quality of student learning" .

Traditional universities present many barriers to the development of more flexible learning. For instance, Robinson et al observe obstacles to open and online learning, many of which are a direct result of the institutional context in which they are attempted: the lack of focus on examinations, slow response times by tutors who are offered little or no reward or incentive to develop online delivery of learning, the ‘not invented here’ syndrome, performance measurement by short term measures of cost against exam results, faculty resistance and fear, lack of support staff (money is invested elsewhere), lack of skills or the incentive to learn them and many other issues. Most universities are not geared around the delivery of courses via alternative means to those which have sustained them for hundreds of years and it is not a practical proposition to attempt to make changes in one part of the system without accommodating that change throughout the system. This is particularly notable in the development of network-based education, with its natural emphasis on self-directed learning. It may be that the whole notion of education as a commodity to be peddled by educational institutions is wrong. Even the most liberal of policies, where (for instance) commitment to lifelong learning and self-directed study are encouraged, are liable to attack on the grounds that the notion of educational delivery implies a system which constrains and disempowers learners by taking away responsibility for what is to be learnt from the learner. Universities thus stand in the way of change, whether intentionally or not. As Illich observes:

Creative, exploratory learning requires peers currently puzzled about the same terms or problems. Large universities make the futile attempt to match them by multiplying their courses, and they generally fail since they are bound to curriculum, course structure, and bureaucratic administration

Unfortunately, this is the education system which not only aspires to lead us to a better life as fulfilled individuals, but which also supplies us with neurosurgeons and air traffic controllers. We need a better way.

2. Conclusions to this chapter

2. : How systems change
1. About this chapter

This chapter will explore a range of related theories which explain aspects of how self-organising systems and their component parts develop. Its purpose is to seek ways that systems are constructed so that some control may be exerted over their behaviour. In so doing I recognise that emergent properties of systems may not (by definition) be designed. However, a watery environment does not usually encourage the development of creatures which breathe air nor do desert plains usually have inhabitants well adapted to climbing trees. In the same way, the eventual goal will be to seek to identify the characteristics of an educating environment, where forms of learning are the likely outcome.

The chapter begins with a look at evolution and the requirements for a system to evolve, giving examples of how it may be applicable to educational systems. In particular, I observe that the kinds of systems considered in the previous chapter owe a lot to evolutionary mechanisms, and attempt to identify ways in which such systems gain their form. With a goal of identifying how structure might develop from the bottom up, I then investigate some of the issues raised by complexity theorists, especially with regard to the ways that order develops, as Kauffman puts it, "for free." The chapter continues with a discussion of the importance of scale and choice of perspective, leading to important conclusions about an appropriate point of view when dissecting the dynamics of educational systems. This will lead to the conclusion that educational systems are driven by many things including learners, their environment and what is to be learnt. However, the most significant part of what shapes educational systems will be seen to be the mediator or manager of learning, a role most commonly embodied in the teacher.

2. Evolution

One of the most powerful mechanisms driving the emergence of order in educational systems is that of evolution, whereby the combination of chance, reproduction, variation and competition results inevitably in a speciated ecosystem. Åm’s example in the previous chapter shows, the kinds of species which develop may not be desirable. My purpose will therefore be to identify the features of an evolutionary system which most drive its development, with the intention of later mirroring those features in a system which evolves to suit its users.

1. A word of caution

The word evolution suffers from many abuses, partly as a result of the redefinition of the word following Darwin (despite the fact that he avoided it himself):

Evolution is to allegory as statues are to birds. It is a convenient platform upon which to deposit badly digested ideas.

When I use the term evolution I will be applying the strong Darwinian sense of the word. I will not be using it as a synonym for change nor for a process of steady improvement.

The principle of evolution can be expressed very simply: if there is reproduction with variation together with competition between the items which reproduce, then those items which are fittest will survive whilst those which are not will die out. The consequences of the theory are immense and profound, playing a large (if not the largest) role in generating the massively diverse order found in nature as well as some of the patterns of history, the spread of ideas and (I suggest) the nature of educational systems.

2. Darwin Lamarck and others

When Darwin published the Origin of Species in 1859 our understanding of the world changed forever. Although the theory was controversial it was quickly and widely accepted despite resistance from those involved in formal religion. Its shock value was caused less by the fact that it dispensed with several predominant notions of the role of a god, but more that it applied a new way of thinking to observed phenomena in the world. Darwinism was the application of a very small number of simple rules to produce a complex system. The power and profundity of this way of seeing the world continues to influence new scientific theory to this day.

Before Darwin proposed the theory which came to be known as evolution it had already been observed that adaptation towards a state of fitness occurs, not least by Charles Darwin’s grandfather, Erasmus Darwin. Dyson claims that "Erasmus Darwin identified the essential principles of natural selection, descent with modification, and other pillars of evolutionary thought." However, the elegance and sheer weight of evidence with which Charles Darwin made his argument make him the one that we now consider the father of evolutionary theory and it is to him that I shall be referring when I speak of ‘Darwin’. Darwin himself gives credit to Lamarck, an early evolutionary protagonist. Lamarck’s evolutionary theory relied on the inheritance of acquired characteristics, citing for instance the growth of muscles in blacksmiths, suggesting that the effects of increased muscle bulk are passed down to the offspring. For Lamarck, improvements to the phenotype (the physical manifestation of form) caused by the behaviour or reactions to the environment of a given organism would be transmitted to its descendants. Although this is manifestly untrue in present populations, Dyson (following on from the work of his father, Freeman Dyson) raises the interesting possibility that replication and reproduction might not have evolved together, and that reproduction without replication might indeed have led to Lamarckian, not Darwinian evolution. If some form of replication occurred without the intercession of genes (for example, we can still see this in the growth of crystals) then changes to the parent could be transmitted to the offspring. It is only when there is an independent underlying pattern of the sort found in genes that the phenotype and genotype lose their strong interdependence. Dyson writes, "We should think twice before dismissing Lamarck because Lamarckian evolution may have taken our cells the first – and most significant- step toward where we stand today." . A similar claim is made by Margulis whose concepts of symbiotic evolution allow for the "infections" of a parent (notably beneficial infections which led to mitochondria and chloroplasts, for instance) to be passed on to the offspring, thus improving the offspring’s’ chances of survival. It is also notable that the concept of memetics developed by Dawkins owes more to Lamarckian mechanisms than Darwinian. A meme is a contagious idea, one which competes in a Darwinian sense with other memes for space in our minds, but there is no underlying pattern or genotype . If a meme evolves, the pattern is entirely visible- the medium is the message. It will become significant to this investigation that the mechanisms of evolution may actually in some ways work better when Lamarckian mechanisms (as opposed to chance mutations) operate, as demonstrated in computer simulation by Ackley & Littman . As Kelly puts it, "Lamarckian evolution produces smarter answers because it is a smarter type of search." Unfortunately, pure Lamarckian evolution is just too complex to operate in the evolution of species. There is no logical reason why there should not be a flow of information from the somatic to the genetic. However, there is no way that the body could identify from the complex system which generates form through genes which genes are responsible for a particular effect. The process of creation of form is far simpler in terms of information flow than the unravelling of how that form can be created, like attempting to extrapolate from the current shape of a cloud the positions of its individual molecules at some moment in the past.

True Lamarckianism has been widely held in disrepute as a mechanism to drive biological evolution, but the Baldwin Effect has been touted as a close approximation of this. Here, learnt behaviours affect the ability of the organism to survive, thus favouring those organisms who can learn a specified behaviour. Those organisms which can more easily learn such behaviours (e.g. through instinct) are better adapted . The idea is updated by Wilson in the form of what he defines as ‘Epigenetic Rules’, whereby culture affects selection which in turn reinforces the rules. For example, poisonous snakes are probably best avoided. This could be a lesson passed on culturally, those who heed it surviving more effectively than those who do not. This in turn eventually leads to a population with a ‘genetic’ fear of snakes . This idea straddles the divide between the concept of evolution and that of self-organisation. A similar notion is found in Bateson’s notion of somatic adaptation and C.H. Waddington’s ‘genetic assimilation’ .

Lamarckian evolution remains significant as it appears to be at least one of the mechanisms underlying, for example, cultural and technological evolution . If any mechanism of evolution underlies systems of education, it is likely to be Lamarckian in nature. Changes and improvements to the phenotype will be passed on through reproduction. For example, good practice in teaching may be mimicked and successful organisational structures will be duplicated. In a sense, perhaps the structures might be seen as genotypical whilst their expression might be the individual instances. However, these structures develop through usage and it is those developments which are passed down. Given that many interconnected systems of this nature will develop, competition will arise. There are, for example, finite ways any individual can teach, and any mode which is selected will have succeeded at the expense of other plausible alternatives. It should be noted that this does not imply that the factors affecting selection of a given mode will correspond to improved learning outcomes. Lamarckian evolution plays a prominent role in CoFIND, the system which I shall be discussing in later chapters.

1. Characteristics of evolving systems
1. Of all the systems views and theories of complexity, evolution has among the longest of pedigrees and perhaps the greatest elegance. If we are to use an evolutionary model of educational systems it is therefore important to understand the effects it brings and how it brings them.
2. Survival of the fittest

The ‘fittest’ in any given system is that which survives in that system. This does not in any sense mean survival of the ‘best’ using terms which relate to anything outside that system. It does mean that death is the greatest teacher, and weaker or simply unluckier parts of the system fail to survive. Their weakness is only in relation to the system. Transferring this idea to an educational system, summative assessments such as unseen examinations, especially those where detailed feedback is withheld from students, are seldom an effective way to teach students anything. They thrive within the system because (amongst other things) of their perceived strengths as an easy means of assessment from a teacher perspective, a reliable means of ensuring that what is assessed is the students’ own work, a powerful extrinsic motivating force (in the absence of intrinsic motivation) and a cost-effective use of centralised resources. These strengths are only in relation to the entire system of which they are a part, not as aids to learning. In a system which does not encourage intrinsic motivation, that certifies learning for other purposes (e.g. helping to get jobs) and that attempts to squeeze as much use as possible out of scarce resources, it is easy to see how summative assessments thrive. None of these features is essential to education, but given their presence in the environment, assessments have a high fitness value. Summative assessments are themselves part of that system and so affect every other part, leading to positive feedback loops which perpetuate the system, much as the trees in a rainforest contribute to the weather patterns that cause the rain which sustains them. This positive feedback is least one of the factors preventing the successful adoption of systems more conducive to learning.

3. An example of the weakness of summative assessment

As a means of consolidating learning, a colleague of recently attempted to enrich her students’ learning experience. As part of a module in systems analysis and design, students were encouraged to implement their designs in the form of a computer program. When the students discovered that this program would not be assessed, for many, work stopped on it. Some students were particularly aggrieved and felt the whole episode to be a waste of time. This transference process which makes secondary goals more important than learning itself is a feature of the summative assessment process. There were two main solutions available to my colleague, either to drop the programming or to make it a part of the assessed work. If it were dropped, then the process of assessment would have defeated a useful learning experience. If it were assessed, then the nature of the exercise would be mutated by the need to meet the assessment criteria, thus preventing the process of discovery and exploration which made the exercise valuable in the first place and defeating a useful learning experience. There are probably other viable solutions such as assessing the process of discovery rather than the products, but whichever method is chosen, the fact of assessment will always shape the process. In this case, like many others in higher education, the issue is complicated further by the nature of the surrounding system. Were my colleague to assess students’ programs, the module would be seen to significantly overlap with other modules which some of the students would be enrolled on, which would either be treated as a ‘bad thing’ (and thus dropped) or provide an unfair advantage to students with modules which overlap. This negative notion of ‘fairness’ is another by-product of the assessment process, driving the mechanistic model of tuition from the bottom up (students demand it) and the top down (teachers and examining bodies believe it to be right). Again, the summative assessment survives, not because it helps learning, but because it and the system which surrounds it have co-evolved to the point at which their mutual interdependencies stamp out new variations. This is also true of most large natural ecologies. A mutation has to be more successful than its competition within a given system. If it is not, however slight the disadvantage, it will not thrive. Even with a slight advantage, the chances of novel forms succeeding are slender. More often than not, the previously established system will correct any changes within it. This is a problem which concerned Darwin and delighted his critics. What is required is a mechanism to allow speciation to not only occur, but to result in change to the system itself. This leads us to another of the cornerstones of Darwinian thought.

4. Speciation

The process of evolution leads to species, a fact of such great interest to Darwin that he named his seminal book on the subject "The Origin of Species." This is of great interest to us if we are to cope with the diversity found in any given group of learners, as well as between different subject matter.

For Darwin, speciation occurs as a natural consequence of descent with modification and competition, hence leading to the survival of the fittest. Contingent occurrences such as continental drift and the isolation of land-masses caused by rising water levels lead to different environments and consequently different adaptations, hence leading to distinct species arising.

Much of what governs the adaptation of biological species occurs as a side-effect of other adaptations, which in their original context are of little consequence but as the environment changes or a species becomes isolated turn out to have the greatest significance. Gould’s analysis of punctuated equilibrium suggests that the pure form of natural selection (strict Darwinism) might be lacking in some respects . What sums up a lot of Gould’s argument is the idea of spandrels- parts of an architectural construction that are simply non-adaptive, side-consequences and architectural by-products of what happens when you put a dome on top of a number of arches, the spaces that are left (figure 2.1).

Figure .1 – detail of Canaletto’s Interior of San Marco showing spandrels

These spaces are not irrelevant:

Under the spandrel principle, you can have a structure that is fit, that works well, that is apt, but was not built by natural selection for its current utility. It may not have been built by natural selection at all. The spandrels are architectural by-products. They were not built by natural selection, but they are used in a wonderful way — to house the evangelists. But you can't say they were adapted to house evangelists; they weren't.

A similar principle may be seen in the evolution of radiator grills on cars. What started as a purely functional requirement to keep the engine cool soon developed into one of the clearest distinguishing features of a car, with huge marketing benefits to the extent that today, where the functional requirement is no longer there and may even adversely affect the aerodynamic efficiency of a car, many manufacturers (e.g. Jaguar, Alfa Romeo, Rolls Royce) still include a radiator grill as a means of creating a visible individual stamp on their otherwise indistinguishable offerings. There are two particularly interesting things about Gould’s analysis in the context of this investigation. Firstly, it emphasises the role of chance in evolving systems. Systems do not necessarily evolve to an optimal condition, but are driven by contingency. The second is that it paves the way for a richer view of self-organisation than that which may be achieved by natural selection alone. Structures can occur for any number of reasons that may or may not join the evolutionary struggle. To distinguish these cases from Darwinian adaptations (which he notes that Darwinians use to represent both the process and the product, as though there were no difference) he has coined the term exaptions. Exaptions provide some of the key to flexibility in evolution. Without them it is unlikely that the chasms left when one species or group of species vacates could be filled. If organisms were perfectly adapted to a co-evolved environment then a major disturbance to that environment such as the extinction of the dinosaurs would have an equal effect on other species, such as mammals. Exaptions might therefore be seen as latent capabilities, able to become adaptations if the conditions arise. These chance chunks of surplus capacity allow entities to take advantage of niches as and when they arise, as in the adaptation of the swim bladder into the lungs of the lung fish, or the adoption of used car tyres as sandal soles in Africa . This is a great advantage of evolution over design, especially design for ‘efficiency.’ As Stewart Brand has observed for buildings, Low Road freedom or High Road flexibility give a much more powerful and efficient way to survive than No Road design Low Road architecturally dull houses which people live in may easily grow and adapt over time, gaining extensions, new facades, being gutted then rebuilt from within, thus surviving changes in use in an almost organic way. Larger, more monumental High Road buildings adapt differently, but still undergo change over many years. On the other hand, elegantly designed buildings which perfectly fit their purpose are in a much more precarious position once that purpose has gone. These are what Brand calls "Magazine Architecture," or No Road designs, perhaps beautiful but immutable and ultimately disposable.

2. A spandrel in the works

Returning to our imagined pseudo-history of the university system, it was perhaps a spandrel caused by the centralisation of resources (itself a necessary adaptation caused by the limited availability of those resources) which paved the way for the lecture as an educational tool. The constraints caused by the small number of libraries and the relatively small number of academic experts in a given field made the university an effective form for bringing about learning. In competition with individual tutors the benefits were clear, not only from an economic perspective but also in terms of quality. As many students were drawn together in a small space, despite its inefficiency from a pedagogic perspective, the lecture was perhaps an exaption, a cost-effective adjunct to the business of teaching and learning, a chance to share ideas with other faculty and students. As student numbers increased, universities and colleges able to process larger numbers thrived at the expense of those that did not. The lecture form proved well suited to this need and was sufficiently effective to achieve the aim of imparting knowledge and (occasionally) enthusiasm and understanding. A spandrel became an adaptation. Natural selection did the rest.

Despite the potential hazards of spandrels, they are a primary motive force in the evolution of educational systems. They are not the only ones, however. In a large population spread over a wide area there is little chance that spandrels could drive the massive diversity which we see around us in biological systems. Small changes would occur eventually perhaps, but larger variations would almost inevitably be swamped in the gene pool. The answer lies in the concept of parcellation.

3. Parcellation drives evolution

Gould identifies a useful piece of information regarding evolving systems:

New species arise in very small populations that become isolated from their parental group at the periphery of the ancestral range. Speciation in these small isolates is very rapid by evolutionary standards....

Major evolutionary change may occur in these small, isolated populations. Favorable genetic variation can quickly spread through them. Moreover, natural selection tends to be intense in geographically marginal areas where the species barely maintains a foothold. In large central populations, on the other hand, favorable conditions spread very slowly, and most change is steadfastly resisted by the well-adapted population. Small changes occur to meet the requirements of slowly altering climates, but major genetic reorganizations almost always take place in small, peripherally isolated populations that form new species.

Calvin makes a similar point about the need for parcellation to encourage evolution , and is expanded upon greatly by Jones and Darwin himself, who devotes the whole of chapter 12 of the Origin of Species to analysing the geographical distribution of species in terms of the formation of niches caused by geological features . The logic is compelling and significant for this thesis. If Darwin, Mayr, Gould, Jones and Calvin are correct, then there is a useful lesson to be drawn for the construction of self-organising evolving learning environments. Scale is important. Without the formation of niches and clustering effects, there is little opportunity for speciation and hence for structure to form. This is taken up in an educational context by Price, who cites a large number of authorities as suggesting that "innovation and learning happens most easily in isolated populations." .

Gould also observes that there is a need to maintain diversity and variability . This in turn is influenced by the amount of space available- "the larger the island, the greater the number of species." If we are to see maximum variability then, we are faced with a contradictory state of affairs, where we want to encourage evolution (hence need small and isolated populations) but also wish to encourage variation (hence need large populations and the large areas they require). This combination of scales is achieved over time by such events as the dropping of ocean levels or a collision of continents.

In existing educational systems parcellation is not difficult to find, although it is often swamped by its larger scale surroundings. For instance, it is rare to find a teacher who does not at some point experiment with different ways of doing things, although such innovations are generally not sufficiently parcellated to have a significant effect on the larger system. Schools, universities and colleges might be seen as larger islands, but their connectivity through channels such as quality assurance mechanisms, conferences, publications and so on, not to mention league tables, word-of-mouth and so on will generally act as negative feedback mechanisms, keeping them within an overall system. However, it is occasionally possible for a university to provide a sufficiently isolated environment for new forms to develop which may then have larger scale effects on the rest of the higher education community. An example might be the model of the OU. In providing for a segment of the community which (for whatever reason) is denied entry to traditional universities, its systems have developed somewhat separately and without a need for much competition with existing institutions. There is so little overlap between the students of the OU and those of more traditional universities that other institutions have provided resources such as teaching staff and building space to assist its development. The OU’s demonstrable success within its parcellated environment is now feeding back ideas into the mainstream university community as well as being reproduced elsewhere such as IGNOU (The Indira Gandhi National Open University in India). It does this at least partly because the Internet has changed the overall ecosystem in which the University exists to more closely resemble that in which the OU has always existed. Parcellated environments delimited largely by country borders are suddenly facing competition from other systems in other countries. British universities are suddenly coming into direct competition with the large online universities of the United States, not to mention a wide range of private institutions from all around the world. The environment has undergone a change where understanding ways of learning at a distance is becoming a necessity in the struggle for survival, and traditional universities are looking for successful models to emulate. The OU, growing and changing within its relatively isolated environment for many years, with ideas, methods and procedures being developed and tested within its own small evolutionary setting, provides a highly developed model, the principles of which are seeping into the worlds of long-established institutions. This is an example of reproduction of successful systems, which will eventually change the nature of the older systems with which they will compete as the landscape that they inhabit changes.

The lesson for the design of systems which will evolve therefore seems to be to encourage parcellation but to allow the occasional collision of ecosystems or the spreading of their areas. This lesson should be heeded by those who wish to centralise the control of education. Without islands of separate development, those variations which may prove useful when the environment changes (as it inevitably will) will not be in place and there is a potential for a catastrophic breakdown, as seen in the demise of the dinosaurs. In the context of this investigation, the parcellation principle will prove useful when designing educational systems which can self-organise.

4. Order for free

The degree of parcellation most likely to enable dynamic change may be predictable. Stuart Kauffman has made many contributions to the discipline of plectics. A notable insight is that of "order for free", order that spontaneously develops at what he describes (after Chris Langton) as the "edge of chaos" . This turns out to be a surprisingly common and predictable phenomenon. Randomly connected networks can achieve self-organised metastasis, although this only happens with certain densities of connection. The best number of connections turns out to average a little more than two for each node of the network.

1. The edge of chaos

Kauffman has coined the term antichaos to represent the limited range of possibilities for complex dynamic systems to fall towards attractors. It appears that any sufficiently connected dynamic system has a tendency to fall into basins of attraction, a sort of metastasis balanced at the edge of chaos. Kauffman warily states that this is as yet merely a strongly supported belief, but it is hugely persuasive. Kauffman gives examples showing the patterns his model predicts in everything from the Cambric explosion to the development of forms of bicycles . The key to this metastasis is the notion of Red Queen and Stalinist regimes of interconnection. Systems which are too rigid (Stalinist) do not adapt to high fitness levels, systems which are too chaotic (the Red Queen effect first noted by Lee Van Valen, always running to stay in the same place) change too much for any stable fitness level to develop. If the landscape is changing constantly then there is no time to adapt. There is an intuitive charm to this. Of all the random ways that connections might happen, it is inevitable that those which survive will reach what Per Bak refers to as a state of "self-organised criticality" . Maintaining a state of order in an educational system without central control will mean providing dynamically connected systems and subsystems, capable of developing so that there is some, but not unlimited interconnectivity.

3. Stigmergy

The word stigmergy comes from two Greek words, stigma (outstanding sign), and ergon (work). It describes instances where agents make behavioural adjustments based on feedback from an emerging pattern caused by that behaviour, mediated through the environment. The word was coined by Grassé in 1959 to refer to systems such as those employed by termites when building mounds. Termites start by dropping mud randomly in a given area. The presence of a mud heap encourages other termites to drop their lumps of mud nearby, with larger mounds being more attractive than smaller ones . This tendency also means that adjacent mounds tend to grow towards each other, thus forming arch structures. This is a classic example of a CAS in action, resulting in the emergence of interlocking arches of great intricacy. The concept of stigmergy is simple to grasp and can play a large role in self-organised behaviour. It also explains the formation of ant-trails. If an ant finds food, it leaves a trail of pheromones on its way back to the nest. When other ants encounter this trail, they are inclined to follow it. If they too find food, they too leave a trail of pheromones back to the nest. As more ants find food, so the scent grows stronger. Eventually the food runs out, no more ants leave a pheromone trail, and the scent dissipates. The ant trail vanishes.

Stigmergy plays a significant role in human dealings. This can be seen in everything from the spontaneous formation of footpaths in a forest, to recessions caused by fluctuations in stock prices, to the large crowds that gather round smaller crowds in a street (an effect well-utilised by professional buskers), to the spread of a piece of software such as a word-processor as a result of the need for compatibility with documents sent by other users of that software. A number of simple individual acts which affect the environment in turn affect other people directly or indirectly. Co-ordinated behaviour arises as a result of indirect communication processes, artefactual communication, if you prefer. In the context of existing educational systems, stigmergic behaviour is rife. For example, the presence of a display of awards in a school hall may encourage those who seek such awards to join the school, as may the publication of exam results. A journal known for publishing high quality papers is more likely to attract high quality submissions. Citations in citation indices lead to more citations. Stigmergy is a powerful concept which is used extensively to drive the CoFIND systems discussed in chapters six and seven.

4. Forests and trees

The dynamics of an ecosystem depend most upon its slower moving parts. This is not to suggest that such an explanation will tell the whole story. For example, zebra and lions move in a constant dance of co-evolution, the evolutionary changes in one determining the chances of survival of the other and its consequent evolution. As predators get faster, prey become more agile or better camouflaged. However, both lions and zebra develop in a context of an environment in which running is an advantage. They would not have developed in the ways they have developed were they living in dense jungle. They would certainly not have developed that way had they lived in the sea, or the Arctic Tundra, or the peaks of rocky mountains. The slower moving parts of the system (of which the geographical location is often among the slowest) dictate the changes of the faster moving parts.

1. A sense of scale

The slowest moving parts of any system really depend upon what kind of scale you are looking at. Brand talks of a hierarchy of ever faster moving parts from roads and natural features to the structural and exterior walls of houses, to the interior structures through to the decorations on those walls and the fixtures and fittings, the evolution of the faster parts determined by the strictures of the slow. Darwin explains much of evolution in terms of the generally slow moving continents that drift and separate and reform as a relentless slow structure determining the changes in their faster changing inhabitants. At any scale within a given hierarchy there will be dependencies which feed up and down the hierarchy, with the fast often (eventually) affecting the slow but with the slow dominating the nature of change. For a microbe, the gut of a fly is a structural feature, a defining control on its possible forms of existence. For a redwood tree, the make-up of the land itself and the weather systems arising from the movements of the sun and moon play a more important structural role. For the plants that grow around the redwood and the creatures that feed on them, the redwood dominates the landscape, literally and figuratively.

2. Stewart Brand and hierarchies of layers

1. An educational hierarchy

Unlike the world of buildings and perhaps more in common with biological ecologies, educational hierarchies are not set in stone. It is common to see different areas of granularity in which different specific entities may occupy a variety of hierarchical levels. For instance, ‘computing’ may be seen in figure 2-2 as a course, a module, a topic or even a lesson, depending on its context. As with most complex systems, multi-scale descriptions are needed to understand the overall system.

There are distinct differences in the hierarchical structure within institutions. However, figure 2.2 represents a fairly typical structure which illustrates the transition from low rate of change to high.

As with Brand’s hierarchy, dynamism and rate of change increases as we drop down the tree. This helps to provide a broad picture of the shaping influences of our existing educational systems. As with Brand’s hierarchy, there is feedback both ways between low and high-stability systems. New resources can shape lesson plans, new lessons can affect modules and so on up the tree. However, the relationship is far looser than Brand’s hierarchy. In a modular system we may see the same modules being a part of different courses, courses may be run by more than one institution and institutions may be part of more than one educational system (for instance, the HND and degree level streams run by many universities). In other words, there is a



Figure .2 – an educational hierarchy

greater degree of multiple inheritance in typical educational systems than might be found in buildings or even biological ecologies. This more flexible hierarchy should in principle allow for greater dynamism within the system. This is the case, but it is not a flexibility which allows the system itself to change, but instead allows it to accommodate all manner of perturbations without disturbing its overall balance. Were there less communication between its constituent parts, it is not improbable that, within separate hierarchies, isolated islands of variegated practice would occur, providing opportunities for new practices to become established as their effects moved up and down the hierarchical trees. It is possible to see a link between our earlier investigation of speciation and our current view of hierarchies. From this perspective, each hierarchy represents a separate ecosystem. With too much connectivity between hierarchies and different hierarchical levels a single system develops rather than several isolated populations, thus preventing the formation of new species of educational system.

1. An example of the influence of the environment of self-organised systems in education

Goodwin writes of the Peckham experiment , which revolved around ‘The Centre’ a club set up in 1926 to improve community health in disadvantaged areas. The experimenters set up a swimming bath, a gymnasium and a theatre, then grouped children by age and allocated times in the facilities for each group. The scheme failed, with few children actually turning up. Success only came when the facilities were made available to the children so that they could turn up individually when they wished. The children did not want to use the facilities in a context of teaching and learning like that found in schools, with instructors in control. To prevent drowning, a degree of order was returned by an ad hoc system of chits issued by staff who would verify that a child was a capable swimmer. A system evolved through a cycle of action and response between the staff and the children. This bottom up approach organised itself far more effectively than the designed system imposed from above, through a combination of enlightened staff and fortuitous building design. Goodwin quotes an account by Sean Creighton ( a member of staff at the Centre):

In my work I was freed from the tiresome restrictions of conventional teaching in schools….

The design of the building was a revelation. That alone was, in my opinion, the master key to the liberation I have referred to. Although primarily designed to facilitate the doctors’ observations, the open plan and windows through which the main activities could be viewed, gave members the unique opportunity to move about the building, to watch others enjoying themselves, and to be tempted to join in and have a go themselves.

Like Gould & Lewontin’s spandrels, an incidental feature designed with another purpose in mind entirely had become one of the main features which allowed the perceived evolution to occur. If we are seeking to create online equivalents of self-organised learning, these lessons would be well-learnt. A system should be sufficiently adaptable to allow users’ interactions to facilitate change in unstructured ways. The openness described by Creighton is a feature relying on critical mass. If there were no one in the pool, no one would be tempted to "join in and have a go themselves". A suitable environment for learning is a pre-requisite for adaptations which allow that learning to occur.

2. Kevin Kelly and chunky bottoms

Given that the environment sets the pace of any given system, the development of that system is not controlled by the slower moving parts, but instead is formed in relation to them. The notion of control has a special meaning in complex systems, as something which is generated by the interactions of its components. Kevin Kelly identifies the importance of developed hierarchies, chunks that coalesce from the bottom up: "no distributed system can survive for long without nested hierarchies of ‘bottom-up’ control" . He likens the problem to the difference between multiplication and division- it is far easier to put chunks together than to divide them. He sums it up:

The Law is concise: Distributed control has to be grown from simple local control. Complexity must be grown from simple systems that already work.

Margulis’s widely accepted theory about the combination of prokaryotic and eukaryotic cells (in the form of mitochondria) is a classic example of this, demonstrating how a major factor for enabling evolution has to be the recombination of already working simpler forms . George Dyson similarly cites his father’s work on the possible ways that molecules capable of replication and entities capable of reproduction might have combined to form the origins of life . As previously observed, Dyson suggests that this would have resulted in evolution according to Lamarckian principles, a point reiterated by Margulis who suggests that, like replication, symbiosis also results in Lamarckian evolution, with the inheritance of acquired characteristics. In such systems, we see examples of where the rules for change themselves change as a result of change. Examples of this can be seen in the numerous taxa which split over evolutionary time. Genetic change causes changes in the rules for reproduction. This resonates with the work of Daniel Dennett, whose concept of a skyhook, a deus ex machina which is introduced by neo-Darwinists to try to explain what they perceive evolution cannot achieve is resolutely denied any part in an evolving system. Instead, Dennett introduces the concept of the "crane," "a subprocess or special feature of a design process that can be demonstrated to permit the local speeding up of the basic, slow process of natural selection, and that can be demonstrated to be itself the predictable (or retrospectively explicable) product of the basic process" . Without such processes, chance alone would have been unlikely to create the complexity which we see around us. Replication was a good idea, so was reproduction. If Dyson was correct, together they provide the mechanism for evolution, each one acting as a crane to lift the other into a new state.

3. Growth by chunking

1. Finding an appropriate sense of scale

You certainly can’t always look at things from someone else’s point of view. For instance, from here that looks like a bucket of water...but from an ant’s point of view it’s a vast ocean, from an elephant’s just a cool drink, and to a fish, of course, it’s home. So you see, the way you see things depends a great deal on where you look at them from

In examining a system, choice of scale is important, and depends upon which subsystem is of interest, from the global to the microscopic:

For what is the Heart, but a Spring; and the Nerves, but so many Strings; and the Joyntes but so many Wheeles, giving motion to the whole Body.... Art goes yet further, imitating that Rationall and most excellent worke of Nature, Man. For by Art is created that great LEVIATHAN called a COMMON-WEALTH or STATE (in latine CIVITAS) which is but an Artificiall Man

The Leviathan described by Thomas Hobbes in his eponymous work is a self-organising system, a group intelligence, a whole which holds together by the complex interactions of its several parts . The world is full of such systems, from the formation of clouds to the dynamics of population growth to the spread of memes in a society . Hobbes’s recognition of the hierarchy of systems of which an individual person is just a component or agent of a larger whole, or an agglomeration of smaller systems, presages an important theme in systems thinking across a wide variety of disciplines: the view of the system is dependent on the observer and that observer’s perspective. Whether we choose to treat a forest as a single entity or just a particular conglomeration of organisms occupying a congruent space and time, whether we choose to look at the behaviour of a single bee or the whole hive, or the ecosystem of which it is a part is to an extent a result of the questions we are asking and the kinds of answers we are seeking. In each instance we are less interested in the matter of which the systems consist and are instead more concerned with the connections between them and the behaviours they exhibit. Our perspective depends upon our needs and interests. As Steve Jones points out,

"Any zoology text claims that there are more kinds of insects than of anything else; but squash a fly and thousands of microbes unknown to science will be squeezed from its gut."

Swift put it charmingly:

Hobbes clearly proves that every creature

Lives in a state of war by nature.

So naturalists observe, a flea

Has smaller fleas that on him prey;

And these have smaller still to bite ‘em,

And so proceed ad infinitum

Swift: Poetry; a Rhapsody

Weinberg is more menacing:

What is a system? As any poet knows, a system is a way of looking at the world. It is a point of view- natural for the poet but terrifying for the scientist!

If we are to guide the development of complex systems, then what is required is an appropriate perspective. As Holland says:

To build a dynamic model we have to select a level of detail that is useful, and then we have to capture the laws of change at that level of detail

We have already seen that most educational systems consist of a hierarchy which can be viewed from a number of different perspectives. As my goal will be to redefine those hierarchies by building from smaller chunks, I shall attempt to identify an appropriate scale to choose when considering how to construct a networked learning environment without too many prior conceptions about current educational systems.

2. Scale in educational systems

1. Deciding the major determinant of structure
1. Learners as determinants of structure

Pines takes the interesting path of inverting the typical top-down faculty-centred view and considering the educational system instead from a student perspective. He notes "Because students naturally cut across departmental lines and make novel connections, student feedback on teaching as well as the direction of research and the internal structure of the university can be especially valuable." As a student centred perspective opens up a far richer realm of complex interactions it is clear that any system we develop with self-organising tendencies should organise from the bottom-up. However, the most dynamic and fast moving of these elements are without a doubt those very learners, as by definition the teacher is attempting to enable them to change. If we seek bottom-up control, we could well find ourselves faced with a similar problem to Kauffmann’s grid of interacting squares. There are simply too many possibilities for interaction and the dynamics of the system may never settle into equilibrium if we allow unlimited interactions between learners to drive the structure of the system.

As we are seeking the slower moving features to identify what shapes the system, we disregard the learner and are therefore left to choose between teachers, context and resources as our main definers of scale.

2. Resources as determinants of structure

There is a whole spectrum of possible structural roles for resources, from those which are necessarily static because they are the object of study (books, plays, buildings, mountains, rivers and populations) to those which are by nature ephemeral or constantly changing (Usenet newsgroups, events reported in newspapers, students’ own work and so on). Depending on the subject and the mode of study, the resources which define learning may be faster changing than the learner or completely static, or anywhere in between.

Teaching may be shaped by resources: it would be hard to teach the history of ideas without a relatively static body of ideas to study. Similarly, most (though by no means all) subject areas consist of a fairly slow-moving body of ideas, arguments and concepts. In the case of, say, the study of a book, content may truly dictate the structure of a learning experience, but even then the manner in which it is approached and assimilated is a variable which is affected by other parts of the system, be they the learners themselves or those who seek to manage that learning.

Resources are promising candidates for key shapers of structure, if not in every situation or context.

3. Context as the determinant of structure

Context plays a vital central role in determining the course of a learning experience. However, in our virtual world of a networked learning environment it is neither a slow moving nor a fast moving part of the system because it is in some sense the system. If we are considering a virtual learning environment, that environment will be made up of the other three constituents, with their relationships to each other determined largely by the way that the environment is constructed and the opportunities that it affords. The context in such an environment is the form that such relationships take, the result of the structuring that has been applied. My goal will be to construct such an environment and in later chapters I will examine the importance of designing that context in far greater detail.

4. Teachers as determinants of structure

1. Conclusions to this chapter

2. : Teacher features

"I cannot teach anybody anything, I can only make them think" Socrates

"You cannot teach a man anything. You can only help him discover it within himself." Galileo Galilei

"I never teach my pupils; I only attempt to provide the conditions in which they can learn." Albert Einstein

"…it’s not just learning things that’s important. It’s learning what to do with what you learn and learning why you learn things at all that matters"

1. About this chapter

Having identified in the previous chapter that the role of the teacher may be seen as central in shaping an educational system, this chapter considers what specific roles teachers actually perform. The chapter starts by considering ways in which learning can occur without an explicit teacher role and the problems that may ensue. It then goes on to consider a small range of instructional theories which provide a framework for understanding the value that teachers add for learners. This is not intended as a comprehensive list of roles, nor is it an unequivocal set of definitions. The intention here is to provide working definitions of how value is added to a system by a teacher, to provide firmer foundations for a self-organising system which takes on some of those roles.

2. Do we really need teachers?
1. Self-teaching

We are happy to talk of self-educated or self-taught people. These are people who have not just absorbed some knowledge along the way but who have in some sense actively pursued and/or created it, who have performed some actions which have led to learning. Learning is more a state of being than a specific task or set of tasks to be performed. As Senge (1990) puts it: "To practice a discipline is to be a lifelong learner. You ‘never arrive’; you spend your life mastering disciplines." If we were to consider the measurement of learning, we might be looking for a demonstrated set of capabilities, not a finite set of tasks that are completed.

A great deal of learning takes place outside of formal settings and institutions. This is an inevitable feature of a rapidly changing world. Papert writes:

As we face a world of ever-accelerating change, it’s no longer possible to have a concept of learning where people in their youth will learn the skills that they will apply through their lives. Learning has to be a continuous matter.

Papert’s solution to this concern is that "what you ought to be learning at school is that you don’t need to be taught in order to learn." . This is echoed by Holt, who claims:

If the child reads only when he wants to, and asks for help only when he feels he needs it, he will work at full capacity, throw himself into the task instead of away from it, and rarely needs help at all

It is a lesson that many of us apparently take to heart, especially Ivan Illich, who observes:

School is an institution built on the axiom that learning is the result of teaching. And institutional wisdom continues to accept this axiom, despite overwhelming evidence to the contrary.

We have all learned most of what we know outside school. Pupils do most of their learning without, and often despite, their teachers.

Research by Alan Tough on self-initiated learning suggested that the average person (young or old) spends approximately two hundred hours each year on some informal learning activity such as learning to sew, learning a new language etc . Tough’s research was limited to a specific culture, and his definition of self-initiated learning is loose, for instance incorporating occasions where the learner actively seeks a teacher or evening class yet excluding ‘learning projects’ which take less than seven hours, but it helps to confirm Illich’s assertion that learning outside an academic institution is the norm, not the exception.

Despite the implications of its name, self-teaching is seldom if ever an isolated activity. If nothing else it will usually be informed by some kind of exemplars which provide goals or starting points for investigation and discovery. More often than not, some element of teaching or training will come from resources such as other people, books, television, audio-cassettes, museums, art galleries or computer-based materials, either in the form of explicit teaching/training material or as primary sources and subject matter from which to learn. Learning is a situated activity, knowledge has a cultural aspect and all education and learning may be thought of as "generic social processes" . This is made explicit in various forms of peer tutoring.

2. Peer tutoring

1. The perils of a chopsticks culture

1. So what do teachers do?

Learning something with the aid of an instructor should, if instruction is effective, be less dangerous or risky or painful than learning on one’s own.

Bruner’s central requirement of a teacher provides us with a useful baseline by which to identify whether or not teaching is taking place, as well as a potential criterion for excellence, but does not provide us with a generative theory. This section will be dealing with theories of instruction which identify what is required for teaching to occur. By abstracting the roles of a teacher I hope to identify requirements for systems which might take on those roles. This section therefore looks at some instructional theories and attempts to extract some common themes.

1. What is a theory of instruction?

1. Abstracting the roles of a teacher

1. A teacher is a subject expert

A merely well- informed man is the most useless bore on God's earth.

The belief that subject expertise is central to the role of a teacher of adults is pervasive. This is carried through to the extent that lecturers are often hired on the basis of their knowledge, not their ability to impart that knowledge. Laurillard rightly pours scorn on this self-defeating strategy , which persists by removing from the teacher the responsibility of student failure. Although perhaps not as central as tradition would have it, subject expertise is none-the-less a significant feature of successful teaching, if only as a means of enabling more central functions, including three out of the four requirements identified by as Bruner. Structure, sequence and feedback are all much simpler tasks if the teacher understands the subject matter.

Teachers do know more of the subject matter and teachers do have responsibility for assisting their students to learn. This means, among other things, not only directing students’ attention to their learning challenge and giving them tools with which to learn on their own, but also helping to remove obstacles that interfere with their learning. And it means understanding that nobody can learn without making mistakes. Your presence and the way you relate to their mistakes can contribute to, or free, your students from fear.

Larsen-Freeman’s suggestion that a teacher is a remover of obstacles and freer from fear is an interesting perspective which echoes Bruner’s justification of instructors as a means to reduce the discomfort of learning which starts this section.

Similarly, Ramsden has a firm conviction that teaching must be situated firmly in a strong grasp of the subject matter, without which it will not be at all clear how best to structure the subject effectively. In fact, he sees it as a subsuming factor:

In good teaching the method used is always secondary to the teacher's aims for student learning and the extent to which the particular strategy actualises the principles of effective instruction within a certain context.

The themes of direction, organisation and goal-setting will arise again in this analysis, but there is no doubt that at some point in the learning experience that a learner will have to confront the subject matter, whether through the teacher (the facilitator of the learning) or through some other teaching resource.

2. A teacher is concerned with communication and the mediation of communication

There is a strong and well founded belief in the world of pedagogic theory that learning is enabled by communication. This is not a new idea. In 1916, Dewey wrote a definitive passage on the subject:

Not only is social life identical with communication, but all communication (and hence all genuine social life) is educative. To be a recipient of a communication is to have an enlarged and changed experience. One shares in what another has thought and felt and in so far, meagrely or amply, has his own attitude modified. Nor is the one who communicates left unaffected. Try the experiment of communicating, with fullness and accuracy, some experience to another, especially if it be somewhat complicated, and you will find your own attitude toward your experience changing; otherwise you resort to expletives and ejaculations. The experience has to be formulated in order to be communicated. To formulate requires getting outside of it, seeing it as another would see it, considering what points of contact it has with the life of another so that it may be got into such form that he can appreciate its meaning. Except in dealing with commonplaces and catch phrases one has to assimilate, imaginatively, something of another's experience in order to tell him intelligently of one's own experience. All communication is like art. It may fairly be said, therefore, that any social arrangement that remains vitally social, or vitally shared, is educative to those who participate in it. Only when it becomes cast in a mold and runs in a routine way does it lose its educative power.

As Bruner says, "the instructional process is essentially social," and is to do with developing ways of understanding through communication. However, communication alone is not the answer to learning. We communicate every day without necessarily learning a great deal in the process, for a variety of reasons such as: we are at the level of expletive, we are making performative utterances, we are repeating platitudes and conventional greetings and so on. For communication to have pedagogical value it needs to be focussed and structured with at least an implicit goal of bringing about learning. A useful framework for that process is provided by constructivism, where the learner is actively involved in the construction of knowledge. Communication is never far from the heart of the constructivist ethos, enabling the interplay of mind and content that is needed to construct knowledge and meaning in the world. For many authors it may be seen that communication plays a pivotal role in allowing learning to take place at all. Pask, for instance, in his theory of teachback establishes that the act of communicating is what brings about learning . Vygotsky’s dialectical structuralism is founded on the interactions with other students and adults who socialise them into the culture . A similar point of view is expanded upon by Muilenberg & Berg , who see the classroom discussion as being capable of promoting thinking of four hierarchically organised and interconnected varieties: critical, higher-order, distributed and constructive thinking. There is little doubt, then, that conversation plays a central role in teaching and learning.

Teachers may be guides of conversation, perhaps in a seminar, tutorial or even in a lecture. In distance learning this role becomes abstracted somewhat. Moore & Kearsley cite Mason as identifying the teacher’s responsibility as falling into three main roles, organisational, social and intellectual:

The organizational role involves managing the conference and providing leadership,. The social role entails establishing and maintaining good relationships and a positive learning environment. The intellectual role involves getting all students to participate and weaving their contributions together

Teachers may themselves be the main source of dialogue. At its most extreme it leads to forms such as Socratic dialogues, where the teacher seeks to persuade the student of a fact or relations between facts by a process of argument. Collins & Berge quote Rohfield and Hiemstra as describing this process as "the responsibility of keeping discussions on track, contribute special knowledge and insights, weave together various discussion threads and course components, and maintaining group harmony." This is an interesting set of roles, especially (pace Moore & Kearsley) those that are involved in ‘weaving, ’ the structuring of discussions and the crossing of borders between discussions. This flexible movement between areas of interest will be seen as crucial to the development of flexible, self-organising systems.

Subsuming the role of discussion facilitator is the role of community builder, connecting learners not only with each other but with the wider community, be it intellectual or social . It is not possible to divorce the practice of teaching from its wider context, as a subsystem in a hugely interconnected array of subsystems. A teacher is one of the primary conduits through which connections are made.

3. A theory of instruction should specify the experiences which implant a predisposition toward learning

Implanting a predisposition toward learning is a complicated feature of teaching, involving not only the more obvious aspects of motivation, but also, following Marton & Ramsden , in an active involvement with the learner, an understanding of what makes a learner curious.

1. Motivation (and demotivation)

As Laurillard points out, one of the few justifications for continuing the traditional lecture format is that it can be inspirational, albeit not very often.

Academics will always defend the value of the ‘inspirational’ lecture, as though this could clinch the argument. But how many inspirational lectures could you reasonably give in a week? How many could a student reasonably absorb? Inspirational lectures are likely to be occasional events.

This view is perhaps a little jaundiced. Few professional performers (actors, musicians, dancers etc) would make a living if they did not fairly consistently inspire their audience, usually on a daily basis. However, we can agree that the consistently inspirational lecturer is a rare beast. Perhaps the most detailed of theories of motivation as it directly relates to a theory of instruction come from Keller and his ARCS model of motivation. Keller bases an entire theory around Attention (need for stimulation & variety), Relevance (desire to satisfy basic motives), Confidence (desire to feel competent and in control), and Satisfaction (desire to feel good about oneself) as the prime motivators in learning. He provides the interesting insight that motivation is not always a good thing:

Motivation follows a curvilinear relationship with performance. As motivation increases, performance increases, but only to an optimal point. Afterward, performance decreases as motivation increases to levels where excessive stress leads to performance decrements.

The word motivation in this context is clearly referring to extrinsic rather than intrinsic stimuli, for it is only in a failure to achieve such goals that stress might arise. Following this theme, it is possible to take the position that every now and then the teacher has to play the role of demotivator. Papert describes a situation where children were so enthused by a project that they were coming in to work during their recreation time and the teacher "had to become a conservative brake to stop them from overdoing" . Although it is likely that the need to stem learning probably indicates a failure in the system, Papert does draw attention to a potential danger: if we are looking towards self-organising systems then learners may be drawn to the areas which enthuse them at the expense of (perhaps vital) other areas, such as those which provide a theoretical foundation for later discoveries. Without a teacher to provide guidance, enthusiasm for parts of the learning experience may result in important foundations being missed.

Motivation is central to many theories of instruction. Bruner states that an instructional theory should specify the experiences which implant a predisposition toward learning . A similar theme can be found in Knowles’s need to foster a co-operative learning climate , Gagné’s’s requirement to elicit performance and Collins et al’s stipulation that a teacher should "involve" students in learning as apprentices .

The need to provide extrinsic motivation may not always be that strong. Alan Tough’s research in the late seventies suggests that around seventy percent of learning takes place outside of an institutional framework and, of the remaining thirty percent, only about five percent occurs with the aid of professional helpers such as teachers, trainers and counsellors . The strong implication is that, where there is a desire to learn, external motivation provided by a teacher is not required.

Teachers have many ways of motivating students, from the threat of exams to the lure of transmitted excitement. McKeachie, quoted in , claims that "one of the major tasks in teaching is not to scare students but rather to nurture their curiosity and use curiosity as a motive for learning." This is reflected in the QAA handout, for UCoSDA- "enthusiasm is a powerful motivator of learning." . When we remember great teachers they tend to be those who enthused us. It is inevitable that in a traditional teaching environment motivation will be low. In a university, class sizes can easily be in the hundreds and the chances of all those students wanting to learn exactly the same things at the same time are very very low. Even with smaller class sizes, the requirements of demonstrating knowledge of a fixed set of facts (and perhaps their relationships and meaning, if the teachers are a bit enlightened) means that it is again certain that some students will be trying to learn about things of no interest to them. It requires a much lower ability to enthuse an audience with an interest in the subject (although the number of boringly delivered papers at most conferences stretches this observation beyond breaking point at times).

4. A theory of instruction should specify ways a body of instruction should be structured for easy assimilation by the learner
1. Resource maker and distiller

Closely allied to the role of subject expert is that of resource maker. From the creation of lesson plans, to accompanying notes, to examination questions, to lectures, to informal discussions, to notes on the board, to whole teach-yourself books, teachers from time immemorial have been seen as sources, condensers and purveyors of knowledge. The mediated transfer of knowledge to the learner cannot be under-rated as a role, and this mediation necessarily involves some creative input. Even in a world where educators as a profession were to disappear, if learning is to continue then there will always be a need for resources from which to learn.

2. Resource gatherer

Whether a teacher is a guide on the side or a sage on the stage, we can infer that a central role he or she plays is the discovery of useful resources for the learner. The teacher acts as a filter, an editor mediating between the tangled complexity of the real world and (by definition) the simpler understanding of the learner. As McKenzie writes:

Good lecturers do tons of extra reading and research into the topic before us. They save us the trouble of doing our own exploration, having ‘turned down the corners’ on the best pages. They synthesize, summarize and report the ‘best parts.’ Knowing their audience, they are able to translate what might otherwise seem foreign, confusing, boring, or overly abstract into a half hour of explanation and illumination. They act somewhat as tour guides.

McKenzie may be a little over optimistic about what a lecturer can achieve in terms of knowing an audience. This may be true in certain cases where the audience is small or of a known state of ignorance (perhaps with no previous experience), but many audiences faced by a typical lecturer are large and diverse. However, in a perfect world the teacher would be an adaptive filter, shaped by the needs of the learner.

3. Resource organiser

One significant part of the role resembling that of a filter is that of a simplifier for representation. For instance, a lecturer may draw a diagram of relationships rather than relying on an original textual description. A diagram of this nature might be found in a textbook, but the lecturer is drawing it for a specific group of learners, adapting resources to be more useful to the cohort.

5. A theory of instruction should specify the most effective sequences to present materials to be learned
1. Definer of boundaries, structurer and sequencer learning

A student may have a desire to study philosophy but may not know even the names of the great philosophers. There are plenty of ways of finding this information, hundreds of books, articles, people who know, Internet discussion fora and so on. The value added by the teacher is to point out the landmarks that are of interest, filter the unnecessary ones and match objectives to the known abilities and understanding of the student. At least, this is what the ideal teacher does. In real life, a teacher with a large group of students provides pointers for an ideal student, an average student or some other general indicator of what a student should be. This may be far removed from the needs of a given individual learner. Knowing the goals is part of the way to knowing how to get there, but it is a lot easier to do this if a structured path is provided.

1. A theory of instruction should specify the nature and pacing of rewards and punishments in the service of learning and teaching.
1. Provider of feedback, assessment and evaluation

Self-directed learners often have difficulties deciding whether learning objectives have been successfully met, and it is an abiding (if often frustrating) part of the role of most teachers to mark work and provide formative and summative assessment. Bruner states that a theory of instruction should specify the nature and pacing of rewards and punishments in the service of learning and teaching . A similar point is recognised by Gagné, Collins et al, Knowles, Laurillard and many other instructional theorists. I have already observed that extrinsic rewards are a poor substitute for self-motivated learning, but the need for feedback (if not rewards and punishments) is a vital element of directed learning.

Assessment can be summative or formative: Formative assessment is generally seen as being preferable from the point of view of the educational system as a whole. From a systems perspective it can act in two distinct ways. In a traditional setting there is a danger that the feedback mechanism is applied like a thermostat. It is feeding back without feeding forward. To work effectively, formative feedback forms not only the student’s view but also that of the assessor. It becomes a part of the evaluation process leading to adaptations in teaching style and content .

Purely summative assessment provides a limited feedback mechanism to the student, but provides an interface between the institution and the world beyond, the rubber-stamping validation that confirms learning has been achieved to those (other than the student) who wish to know. Its role in the system is one that acts as a channel between the educational system and the world in which it resides. In principle, if those who are qualified are poorly adapted to the world of employment then feedback from the employers will affect potential students which will in turn affect intake figures to the awarding institution. In practice, the world is a slightly different place. The relationship between learning, assessment success, employability and institution status is richly interwoven with other factors. Choice of institution may be affected by location or history, student success related to prior ability and so on. Employers may have preconceptions about institutions and may themselves be uncritical former students, so that the complex whole achieves a kind of solonic wave of stability. At the University of Delhi, for example, computer courses are developed in conjunction with a large local employer, but the representatives of that employer are themselves former students, whose preconceptions of what constitutes a useful course is based upon their own schooling. Thus, it remains an entry requirement to know about the Brownian motion of gases and nuclear fission, despite their apparent minimal relevance to the subject of computing.

The evaluative process does not end with assessment. Evaluation of learning resources, of learners, of the learning of learners, of teaching, of environments, of learning technologies and so on is a major role played by the teacher. Moore and Kearsley identify direct questions, assignments, quizzes, polls and questionnaires as being effective means of getting this kind of feedback . It is also reflected in the principles of reflective teaching espoused by Marton and Ramsden .

2. An holistic view

Collins & Berge represent design of the student learning experience as something to be striven for in contrast to just the provision of content . The objective is to provide scaffolding to allow students sufficient latitude and freedom to refine questions and search for their own answers from multiple perspectives. This role cuts across or straddles many others and is so non-specific that it could largely be construed as being synonymous with ‘teaching’ itself. However, I would like to retain a sense of the phrase inasmuch as it provides an holistic view of the activities. Although this process of analysis has considered a group of more or less distinct roles, like most complex adaptive systems teaching appears to be more than the sum of its parts, not a simple agglomeration but a system of interacting elements.

3. Other things a teacher might do

1. Conclusions to this chapter

2. : Hillocks in the stuff swamp
1. About this chapter

The previous chapters have discussed the interconnected nature of educational systems, identified a fair selection of the features which distinguish guided learning from unguided learning and considered some of the ways in which emergent behaviour can arise in complex systems. This chapter looks at ways in which emergent guidance might arise as well as how it might be hindered within network-supported learning. It is an investigation of ways in which self-organisation takes place in networked learning environments without explicit support from software. To begin with it will address some of the problems of searching and ways in which structure emerges when seeking resources on the web. I then report in some detail on a few small-scale instances of asynchronous discussion groups that have been used in my teaching, with the aim of discovering their strengths and shortcomings in allowing self-organisation to occur.

1. Teachers are bad for learning

In the first two chapters of this thesis I suggested that the nature of educational systems affects their subjects, sometimes negatively. In many cases, this occurs because of the hierarchical structures that pervade educational systems, leading to rigidity, lack of adaptation to learners’ needs and unwanted artefacts which are antithetical to effective learning, such as examinations and lectures. I suggested that the most significant layer in this hierarchy is filled by those who teach. Despite teachers’ roles in perpetuating potentially harmful systems, it became clear in the previous chapter that they are important enhancers of the learning process.

We are faced with the potential quandary that teachers, whilst apparently a boon to the process of learning, are typically major contributors to systems which are detrimental to learning.

There are a few possible solutions to this problem. For example, a significant amount of educational literature is concerned with ways in which teachers can and should become "guides on the side," thus reducing their hierarchical roles. With sufficient care, these roles might genuinely shift to the side within the hierarchy. One-to-one tuition, which at least allows the possibility that a less hierarchical relationship might develop, is almost universally hailed as a most effective way to assist learning, but economic realities make this the exception rather than the rule. An alternative that will be explored for most of the rest of this thesis is to consider ways that the roles of teachers might be replaced, or at least extensively modified, by making use of the combined intelligence and knowledge of a group of learners. This is not a particularly radical suggestion. We can see the process in many conventional learning settings, such as the traditional seminar or the academic conference, where peers share their research and learning and, through a process of discussion and debate, help each other to learn.

The concept of a teacher-less future will be presented as an alternative option, not as a cure. It is not necessarily a better idea than the status quo, and may itself lead to unexpected emergent unpleasantness. We are in the world of complex adaptive systems, where it is often hard to predict behaviour based on simple rules.

The growth of the Internet allows a range of different forms of interaction and means of information retrieval and knowledge sharing not afforded by traditional situated methods of education. Many of these forms make explicit their utility as systems to aid collaboration. Network-based learning is therefore a fruitful area of investigation when considering the potential for self-organisation in learners.

2. Finding good stuff

First, what's there is stuff: partly information, partly pure nonsense--and it's not always easy to distinguish the two. Second, it's not a superhighway, it's a swamp, albeit a swamp with many remarkable hillocks of well-organized, first-rate data and information.

In universities the hiring policy for lecturers has often been far more concerned with their perceived knowledge than with their teaching abilities. It is assumed that those who possess knowledge will somehow be able to transmit it to their students. If equivalent knowledge could somehow be transmitted in another way, then it is tempting to speculate that, at least from the point of view of those who specify such selection criteria, little would be lost if lecturers were removed entirely from the system. This is, of course, silly. I have already gone to some lengths to dispel the notion that knowledge transmission is the sole function of a teacher, but the passing on of knowledge and the ability to respond to questions of subject matter are important roles which might in principle be achieved without a traditional lecturer.

The Internet provides massive diversity and a huge range of people and resources. There is no shortage of information, with most subjects being served by thousands, sometimes millions of pages, discussion groups and so on. The main problems of utilising this plethora of knowledge revolve around discovery and structuring of relevant and reliable resources and people . These problems divide into a number of distinct areas, including those of lack of shared categories and values, inconsistent ontologies (in the computer-science rather than philosophical usage), inadequate indexing, lack of semantic metadata, and the sheer volume of the web, currently estimated at several billion pages on around seventy-one million Web sites and growing . This is not including the indefinitely large number of dynamically generated pages and the hundreds of constantly changing messages on a large percentage of around fifty thousand newsgroups and an even more enormous number of mailing lists and ezines. Even were all of these problems to be solved, learners would still be faced with major difficulties in identifying appropriate resources for specific learning needs. It is the nature of learners that they are unlikely to know where to go next and indeed it is the logical outcome of learning that prevents this from ever being an easy problem to solve. Learning changes the learner. This is a necessary feature of any system that learns. As Kosko puts it, "learning changes an information medium" . Future needs are generally hard to predict prior to that learning taking place.

1. Brute force and ignorance

I have identified that resources are a central feature of abstract educational systems. In a traditional educational system, these are usually discovered and explored with help from a teacher, but the example of self-taught learners shows that this need not be so. Learners may make use of self-teaching books, for example, or each other to discover a subject. Networks provide a potential to share the burden of discovery efficiently. One individual may discover or create a small number of resources. Many users will be more likely to discover or create many resources. Computer networks allow these discoveries and creations to be shared. Diversity and conflict between ideas that arise generate a knowledge ecology. Malhotra writes:

Just as natural ecologies thrive based on species diversity, knowledge ecology thrives on diversity of knowledge. Such diversity rests in co-operative competition: the various knowledge nodes collaborate as well as compete based on their differentiating characteristics.

Knowledge ecology treats knowledge creation as a dynamic evolutionary process in which knowledge gets created and recreated in various contexts and at various points in time

This chapter will be looking at systems which allow this ‘co-operative competition’ to occur. Such systems will incorporate methods for communication and sharing, as well as providing a rich diversity of knowledge to be drawn upon. Networks, especially large ones like the Internet, provide these special opportunities for mediated communication and resource sharing. They allow the creation of various different 'landscapes' for interactions to occur, as a means of structuring and controlling the ways in which users react with other users and learning resources. In the next chapter I will explore many fascinating examples of systems which make use of and enhance this effect, but there are plenty of instances where these landscapes are quite perceptible without the aid of special software. The rest of this chapter looks at how such landscapes occur and the limitations of a relatively ‘raw’ approach to identifying them.

1. The Small World Web

Brown and Duguid have observed that all documents are in some sense related, whether on the Internet or not:

From turned down pages, to notes on a dust jacket, to academic essays, to fan zines, to direct quotations and indirect allusions, to stories lifted for future retelling without attribution, we are always commenting on texts, which continually intertwine in a process grandly known as "intertextuality." Documents are not, then, independent. Like biological organisms, every document is always related to some other.

The connected nature of the Internet reifies these connections and multiplies them inordinately. The notion of a "small world" topology refers to a network with highly clustered nodes. The classic example is the "Six Degrees of Kevin Bacon" game, wherein the object is to find the shortest path from any actor to Kevin Bacon. An actor with a Bacon Number of one has appeared in a film with Kevin Bacon, an actor with a Bacon Number of two has appeared in a film with an actor who has appeared in a film with Kevin Bacon and so on. It turns out that it is extremely difficult to find actors who have appeared in films with a Bacon Number of more than four . Adamic has analysed the Web to find that it too has a small world topology. It is this feature that gives the Google search engine its power, as it identifies the degree of linking to a given page from other pages. Giles reports on research at the University of Notre-Dame which suggests that, of the billion or so pages available on the web, the average distance between two randomly selected pages is in the region of nineteen clicks apart . This clustering phenomenon suggests that even without the aid of adaptive hypertext there is at least a certain amount of self-organisation occurring on the web. Adamic reports on a strongly connected group of sites relating to the pro life and pro choice communities by examining cross-links within them and is able to conclude that the community of pro lifers is more tightly knit and better organised than that of the pro choicers. Lempel & Moran also investigate and confirm the same range of clusters which exhibit the TCK (Tightly Knit Community) effect as a means of proving their own LHA (Latent Human Annotation) algorithms. Simply through the independent linking of pages, the Web in its ‘raw’ state might provide at least the beginnings of the organisation required to structure knowledge.

2. Web visualisation systems

Kleinberg and others have capitalised on the inherent patterns within the web, leading to systems such as CLEVER , SALSA and Google (http://www.google.com). These systems iteratively mine the Web for clusters, based on an analysis of in-degree and out-degree in hubs and authorities. Information that Kleinberg describes as latent human annotation (LHA) is condensed and purified, resulting in emergent patterns indicating popular and probably relevant sites of interest, given a few keywords. Although such sites may be authoritative and even good, LHA- and citation-based systems are not particularly likely to be especially helpful to learners. This is because the kind of structures which are found are not of a form which corresponds to the kinds of structure I have identified as being generated by teachers. Although a system such as Google will provide us with implicit metadata (discovered resources are displayed in order of relevance, hence we are looking at implicit metadata describing degree of relevance) it does not capture the context in which the data are to be used. It does not tell us whether what we visit will be useful to us at our present stage of development, whether we should look at something else first, even whether the information we are presented with is within our frame of reference. Unlike a teacher, the only significant value criterion used by such a system to describe a resource is its popularity. In a sense, we might treat this as a single dimensional system, where value is plotted on a straight line.

An extension of the principle of clustering by indegree and outdegree as used by search engines such as Google and Raging is taken further by visualisation systems based on data mining, such as Document Explorer, Harmony and Narcissus . These systems provide a visual indication of the relative popularity of a resource and the degree of linking to other such resources, thus identifying emergent structure based on the principle of LHA. By providing a visual representation of these links, the inherent structure of clustering in a group of Web pages is highlighted, leading to the emergence of order based upon the independent decisions of individual authors to provide links from their Web pages. This is a powerful example of self-organisation, although it does not provide a path through the resources nor give any clue as to why the pages might be considered popular: there is no semantic linking here. It is possible that such authorities will turn out to be sites that have been advertised for money rather than placed there because of their intrinsic value. Even if means of avoiding such hubs are implemented (and Google uses a variety of strategies to do so), there is still no indication of the pedagogic value to a learner. We can identify self-organisation but not to instructional ends.

Self-organisation of some sort can even be seen in the interactions of a simple discussion forum.

3. Discussion fora, flora and fauna

In looking at the ways that networks can assist the development of self-organising learning environments we do not have to stray into the realms of the exotic or unusual. Take for example this extract from the late Tom Creed’s seminal article Extending the Classroom Walls Electronically:

... my students E-mail me their written homework assignments two hours before class. Not only does this allow for incubation for them, but it provides a unique and invaluable classroom assessment technique for me. I can read my students' submissions before class, which allows me to assess what they know BEFORE class begins--I go to class having a good idea what they know and where the problems with their understanding are. This allows me to skip material that they seem to know well already, and concentrate on the areas where there are widespread weaknesses. In addition, I read their Make Up Your Own questions, so I can address the topics that they have already identified as being of interest to them. Similarly, the ongoing discussion on our electronic conference gives me vast insight into how they are viewing the more controversial areas of the course. Our discussions are enhanced since we all have an enlarged shared cognitive set--we all know what each other is thinking about the topic.

In this simple system we see many of the hallmarks of a classic self-organising system. One agent (the student) affects another (the teacher) which adapts its behaviour and thereby affects the original agent and so on in a rich cycle of adaptation and self-adjustment. In this example the teacher is an important part of this system and plays a significantly different role from that of the students. Given our goal of self-organised learning, this is too hierarchical, relying on a single agent to provide structure. What I will be aiming for is to see such significant roles of the teacher either as an emergent property of a system or as a set of roles that form from the randomly differentiated learners who make up the system.

A self-organised learning environment is not unusual, as least in part. The concept of the seminar, with is structured presentation/discussion approach encourages the development of learning simply as a result of the interactions of students, perhaps but by no means always with some teacher-mediation.

Murphy et al discuss a number of uses of CMC (Computer-Mediated Communication) technology to provide support for self-organising groups of learners. Technologies used varied from BSCW (Basic Support for Co-operative Working, at http://www.bscw.de) to MOOs to FirstClass, providing a variety of integrated and less-integrated tools for asynchronous and synchronous discussion. In each of the instances reported, the students were not exclusively taught using CMC and (more significantly from the perspective of this thesis) teachers played a strong role in providing a structure for self-organised learning to occur. Teachers provided tasks, split students into groups, gave them roles to perform, and scheduled opportunities for meetings. However, there were clearly some aspects of the various systems which were organised by the students, if not self-organising in the strong sense that I am seeking. Students moderated discussions, performed some scheduling tasks, designed evaluation criteria and went through an iterative process of design and evaluation to produce project outputs. Murphy et al describe the advantages and disadvantages in terms of collaboration, relevance, learner control and technological preparation . Although some of their conclusions are a little trite ("adds realism to the working process", "synergy results from two minds working together") there are some important insights. They note the problems experienced by some students with leadership roles and (conversely) the problems of ‘too many chiefs, not enough Indians.’ A related problem is identified as the intimidation of the less well informed or skilled. Social roles such as defusing such intimidation and providing some form to discussions are usually performed by teachers and are important to the smooth running of a designed system. If we are to build a self-organising system then these issues need to be taken into account. On the positive side, Murphy et al talk of the benefits of shared workloads, multiple approaches and diversity, control of time and pace and advantages of students learning from each other. They conclude that "group formation and group dynamics are critical to the success of collaborative learning and peer-directed activities." Establishing the means to allow social networks to develop without losing the periphery is definitely a concern, although it is hard to tell from this study the extent to which the teacher’s control and need to retrofit the results into an existing system causes the need for the particular kind of roles that Murphy et al see. A similar problem affects my own experiments. Subsystems of learning cannot be divorced from the systems of which they are a part. It is quite difficult to identify the effects of other parts of the system on the development of new subsystems.

4. My own experiments

Messages are themselves a form of pattern and organization. Indeed, it is possible to treat sets of messages as having entropy like sets of states in the external world. Just as entropy is a measure of disorganization, the information carried by a set of messages is a measure of organization. In fact, it is possible to interpret the information carried by a message as essentially the negative of its entropy, and the negative logarithm of its probability. That is, the more probable the message, the less information it gives. Clichés, for example, are less illuminating than great poems.

Weiner’s description of messages as a form of pattern suggests that we may be able to treat messages and their interactions as elements in a self-organised hierarchy. It might be possible to consider the interactions of a newsgroup in support of a set of learning objectives as a self-organising evolving system. The following section builds on work first described in .

One of the most useful types of resource for groups of learners with shared interests is the asynchronous discussion forum, exemplified by Usenet News and Webboards (Web-based discussion fora).

1. Methodology

The three cases that are about to be described make use of the students’ own comments either in the forum or in reflective documents about the exercises. No structure to these comments was imposed and no questions were asked to elicit the responses, which may thus be seen as fairly freely given. The purpose of this study is to interpret, not quantify. This is not an objective study, as (for example) the various agendas of the students themselves may well have influenced their behaviour, perhaps perceiving that they might achieve better marks if they pleased the tutor with meta-comments on the process, or maybe merely wishing to better organise their thoughts on the learning achieved. The comments are reproduced as originally written, without any corrections to spelling, punctuation or grammar.

1. The first study: raw fora

To investigate the power of newsgroups to facilitate self-organised learning, I created an exercise for a group of level one computing students of very mixed abilities . The students were set the task of exploring a topical subject at the time, that of net computers (NCs) versus personal computers (PCs). This was selected as being not only topical but also a ‘wicked’ problem. Wicked problems are a "class of social systems problems which are ill-formulated, where the information is confusing, where there are many clients and decision-makers with conflicting values, and where the ramifications in the whole system are thoroughly confusing " . The great advantage of setting such a problem is that there is unlimited scope for argument and discussion. The students were required to use this newsgroup for discussions and asked not to discuss the issues using other means.

The group of 102 students on a basic computer skills module known as Personal Computer Concepts were given no background information and none had any significant prior knowledge of this topic, although some were a lot more experienced with and capable of using computers than others. The exercise took the form of an assessed assignment and was set to last for two weeks. I explicitly avoided participation in the discussion myself, so that the students knew that they were on their own. Thirty percent of the marks for the assignment were given for level of participation in the discussion, with the remainder going to reflective summaries of the discussion presented as Web pages. None of the students were expert news users, though all had taken part in a simple introductory tutorial.

1. Results

An average of around 2.5 messages per student were posted, with a maximum of twenty from one student and five students not contributing at all. Two hundred and fifty messages were posted in total.

Most of the students believed that they learnt something and it was clear that the newsgroup had played an important positive role in this learning experience. The students’ reflective summaries were mostly very positive with regard to the conversational processes of learning engendered through the use of the newsgroup. For example, these comments from various students demonstrate their growing awareness aided by others:

"I found the newsgroup to be beneficial in both learning facts and gaining a wide variety of views that I had not considered. This is the first time that I have participated in a discussion of this sort and I was pleased to find how well it worked in providing fact and opinions and then prompting relevant discussion."

"There are various opinions on this Newsgroup, some of which I agree with and others that I do not. The main benefit of this group, to me, was how items have been discussed that I hadn’t realised or thought of."

"I felt that using the newsgroup for this exercise was a good idea as it was very useful to see volumes of opinions on such things as NET PCs, Web TV etc. Before forming one of my own."

Coming from the students themselves, the level of technical knowledge was well adapted to their own level of understanding, as demonstrated by these two comments:

"The discussions that have occurred as a result of the ... newsgroup have been very productive. To begin with those students who had difficulty defining the topics involved have been able to seek the help of their peers. This has been beneficial to everybody mainly because the explanations that have been posted have been in the answerer’s own words, not jargon from a book or magazine."

"The discussion was good because it made it easier for people to understand what was going on as it was in people’s own words."

Movement from a lesser to greater knowledge was apparent with the passage of time, with a limited amount of structure to that knowledge imposed by the hierarchical threading of messages. Broadly speaking, messages towards the end of the thread conveyed more refined knowledge than those at the beginning, although this was not always clear nor consistently the case.

Although there was a modicum of evidence supporting the self-organisation of learning, there were some notable problems that show a failure of the newsgroup discussion to adequately organise itself to improve the learning experience of the students. In particular, I identified trust & reliability, losing the thread, lack of differentiation between messages and hostile postings as being particular issues, which I discuss at length below.

2. Trust and reliability

In keeping with Powell’s comments on tutorless groups there was some uncertainty expressed about the reliability of messages posted. Several students commented on the problem of trust in the authority of their peers. One student noted:

"Having scanned through the messages ‘posted’ on the newsgroup – I realised that there were a great number of people, like me, with very little knowledge on the subject in question- but there were a few prolific ‘posters’ who seemed to have a good technical understanding of what was going on. Unfortunately, despite the latter group, a lot of basic questions posed by students remained unanswered. Hence, I found it difficult to build up a solid picture from all the little snippets."

Several interesting issues are raised by these comments. Although a lack of trust is indicated, there is also a suggestion that some students became recognised authorities very quickly. This is self-reinforcing mechanism, as Heylighen observes, pointing to simulations performed by Gaines where "individuals who were successful in solving a particular type of problem are more likely to get problems of that type delegated to them, and thus will develop a growing expertise in the domain." This view is strengthened by the fact that those perceived as having authority were also the more prolific posters of the group. This feature can often be seen in mailing lists and newsgroups alike. It is reminiscent of the notion of ‘key species’ alluded to by Watson , those "irreplaceable species and functional types" Watson claims are analysed by Grimes as being more important to a functioning ecosystem than simple diversity. Whether or not this notion is correct, it is clear that a few posters had a disproportionate effect on the functioning of the group as a whole and that this is a pattern seen in most online discussions. Without the incentive provided by marks it is highly improbable that more than a few of the students would have participated at all. In large, public groups, this is endemic.

"A recent survey by a computer consulting firm in Chicago found that 98 per cent of the visitors to large sites with open forums - from AOL and MSN to sites like Slashdot - never submit ideas or articles and never post opinions or participate in arguments. "

The amount of lurking (reading without posting) within an online discussion varies from group to group. For example, Preece reports that in health support groups lurkers may comprise less than half the total of "participants" in a discussion group. It is far from clear that lurking is necessarily an evil practice, and there is some doubt about research methodologies which have tended to look at groups over a limited period to identify lurkers. I am a subscriber to some mailing lists to which I contribute once or twice a year, which I judge to be more than adequate in those contexts, but I would be considered a lurker in studies such as that of Sproull & Faraj which looked at contributions to a discussion over a single month to find that 80% of people subscribed to the discussion were lurkers. In my quiet periods I am still actively immersed in discussions which I follow closely.

The problem of building up a picture from small snippets is a more pernicious danger than lack of trust or (worse still) inappropriately placed trust. One of the key roles that I have identified for a teacher is that of an organiser of information. To recap, this organisation involves (in an ideal world) selection and sequencing of material appropriate to a learner’s needs. Without this role, the snippets of information to which the student refers may remain disjointed and incomplete, at least in relation to the topic as defined by the teacher. If the choice of learnt material were in the hands of the learners, there would perhaps be a very different definition of incompleteness, in terms of the needs of the learners rather than the prescriptions of the perceived ‘expert’. Within the context of the exercise, however, there was an apparent movement from a state of little knowledge to a state of greater knowledge throughout the experiment, but that knowledge belonged mainly to individuals. From a broader systems perspective there was a lot of barely differentiated information but little consensual group knowledge.

3. Losing the thread

There was a certain amount of drift from the set topic, which might be seen as an indicator that the guiding role of a teacher is valuable. In particular, a whole thread formed about the future of the floppy disk that was at best tangential to the main discussion. As an evolving method of learning this seems appropriate and potentially rewarding, with the possibility of allowing students to follow their own interests. However, the main reason for this occurrence seemed to be the students' limited knowledge of the subject, leading them to feel their way into adjacent territories. The species did not survive for long, self-organisation setting in when it was noted that this was not a fruitful line of discussion. That said, this provides some evidence that mutation and variation can occur quite unexpectedly and it was perhaps just bad luck that this particular variation was unsuccessful.

4. Undifferentiated volume of messages

Mostly through the students’ lack of experience, few threads developed further than three or four levels deep, with a fair number of posted replies being sent as the start of new threads. The lesson here seems to be mainly that any technology has a learning curve. Given time, learners would probably come to grips with the possibilities afforded by the technology, but two weeks was apparently not long enough for this to happen. The lesson we need to draw from this when building our own self-organising environments is that the technical hurdles of using it should be kept to a minimum. Perhaps Kelly’s maxim that systems should grow by chunking , building larger from smaller systems might be appropriate. Revolutionary systems are inherently harder to master than those with which learners are familiar.

5. Hostile postings

One student became quite aggressive in his postings. Within the controlled environment of this exercise I was able to stop this with a single email message, but it could become a significant problem in a self-organised environment. Flaming (the posting of aggressive messages) is a common problem in newsgroups and mailing lists in general and eventually, like a forest fire, will usually die out of its own accord without destroying an entire newsgroup. Indeed, in moderation it probably provides a useful means of feedback control, preventing careless or unintelligent/uninformed postings. However, the casualties along the way are those who are dissuaded from posting for fear of reprisal. This is a clear instance of a self-organising evolutionary environment, but it does not adapt to better suit the learners. Instead, it sets up a predator/prey relationship in which only the mean or resilient survive. There is a lesson to be drawn about the role of a teacher as moderator, which needs to be incorporated into any self-organising environment that we may build. It might also be worthwhile to investigate ways of making the effects of flaming less traumatic to its victims.

6. Technical problems

The students mostly used Netscape News (version 3) which had a number of ‘features’ thankfully removed in version 4 and above. In particular, the default settings limited the number of messages downloaded to one hundred at a time. This meant that some students posted messages then thought that they were lost.

Due to security restrictions on the University of Brighton News server, students attempting to gain access from home were limited to using Pine or Tin via Telnet, which were universally detested to the point that no one used them. Technical and interface problems are a feature of any interaction with a computer system. A system with the soundest theoretical basis will fail if it cannot be used. Even a well-designed and reliable system requires users to learn different ways of seeing and behaving.

7. Conclusions to the first study

Although many useful lessons were learnt (in at least two senses of the word) the newsgroup showed itself to provide at best limited support for self-organisation. My own involvement was a major factor in shaping, but I was not seeking autopoiesis (the system pulling itself up by its own bootstraps), merely for signs of self-organisation within the strictures of the environment. As Churchill put it in a speech made in the House of Commons on October 28, 1944, "we shape our dwellings and afterwards our dwellings shape us." . The biggest obvious problems were lack of experience with the medium, trust & reliability, and losing the thread due to the myriad of entangled interactions.

2. The second study: Adding structure.
1. Lessons learnt

I identified a range of issues arising from the first study. Some of these could be traced to lack of knowledge or experience on the part of the students, especially technical problems, inadequate use of threads and lack of knowledge of netiquette. These problems could be solved by providing experience and traditional tuition to students in the use of the medium. Reliability and losing the thread could not easily be resolved this way, however. Reliability was an issue that I was not yet ready to deal with, but I devised a further system to enable students to learn which attempted to deal with the organisation problem by imposing a structure on the discussion.

The second study took place some five months later than the first. It involved a group of 32 students enrolled on a one year MSc Information Systems conversion course, studying network technologies. The students were similarly lacking in familiarity with computers as the first group, although in other ways (especially in terms of commitment and prior education) they were quite different. The task was also significantly different, though again involved the discussion of wicked problems. For the purposes of this piece of research these differences did not seem significant as its purpose, like all the research I have carried out for this thesis, was to identify potential avenues of development, not to answer questions which had already been posed.

I had ensured that the group was much more familiar with the technology than the earlier group, to which end the students were given a tutorial at the beginning of the course, followed by a two-hour tutorial with exercises just before the assignment. From the start of the year they had been repeatedly encouraged to use the newsgroup for course-related communication and it had been well used by around half of the group, with evidence that nearly all the remainder spent time lurking. This was strongly encouraged as last minute changes to timetables and other such announcements were posted to the newsgroup, sometimes as the only source of such information. As it was also being used to aid administration (events and new timetables were posted there throughout the course) most of the students were familiar with reading Usenet news and many had used it to seek assistance.

2. The assignment

1. Structure and guidance

By creating a distinct structure and explaining what I considered to be good practice, I had placed a number of constraints, some of which were clearly stultifying. One student wrote:

"We receive guidance in how we are expected to contribute to the exercise which could be seen to have shaped the way we went about it. Could we not have collectively decided that there was another way we wished to go with this?"

Despite trying to take a back seat and to enable self-organisation to occur as far as possible, I had already made my mark by imposing the structure and criteria for excellence of the exercise. One of the previously identified roles of a teacher is to establish the scope and then allow students to explore the realms that lie within that scope, but it was difficult to decide the boundaries between interference and appropriate guidance. It is important to make work as relevant to the students’ experience as possible and to teach them what they need to learn when they need to learn it, or at least to make it clear why they are doing something. A student wrote:

"How about some more spontaneous threads that really reflect our real interests/concerns? "

This is an important issue, relevant to William Glasser’s Control Theory, of which On Scope Associates write, "If students are not motivated to do schoolwork, it is because they understand schoolwork as unconnected to their basic needs" . This view of intrinsic motivation arising from a genuine interest in the subject is a vindication of my quest for emergent education, based on learners, not the demands of teachers and the systems that surround them.

Perhaps the neatest summary of the (not quite fulfilled) potential of this system comes from a student who wrote:

"One of the potential advantages of this form of exercise is that it brings the possibility (for the first time on this course? with the possible exception of self-organised pre-exam revision workshops) for the group to take charge of an exercise and pool its not inconsiderable collective intellect."

It is clear that an unadulterated newsgroup, particularly where the footprints of the teacher are so visible, only provides partial support for this process. There is a strong need for more support for structure and control emanating from the participants themselves.

2. The outside world intrudes

1. The third study: adding stigmergy

The imposition of structure seemed to have had a moderately positive effect on the growth of discussions, but the issue of trust still remained and it was clear that the basic hierarchies of the discussions were not adequate to provide the kinds of self-organisation based on students’ needs rather than the structural Stalinism of the environment itself.

When setting the same exercise for the following year’s intake of forty-four MSc students, I dropped the use of the newsgroup in favour of a home-grown solution, largely to examine the effects of a simple form of collaborative filtering on the ability of the system to self-organise. A subsidiary goal was to improve accessibility to students working from home, an option available to over ninety percent of the cohort. The system I created was a Web application written in Microsoft’s ASP. It incorporated most of the features found in newsgroups, with the addition of a search mechanism and a simple rating system- a button at the foot of each message saying, "I found this useful." The results of pressing this button were fed back to the users by a simple numerical indicator showing the number of votes for each message in the message listings (figure 4.1).



Figure .1 – messages with stigmergic ratings

In this way I was attempting to provide a form of positive feedback which would lead to stigmergy. The forum was also divided more firmly into themes, selectable via a drop-down list, within which threads were allowed to develop. There was not enough time to develop a more sophisticated structure than the simple hierarchy.

In most other ways the exercise was similar to that of the previous year, with the significant change that instead of a "post useful URLs here" thread I provided a full-blown collaboratively filtered bookmark mechanism which was also used to post finished work (CoFIND 1) which is discussed in chapters 6 and 7.

1. The meta-discussion

As with the previous year’s news discussion, information overload was cited as a significant problem, though only by two group members and then almost frivolously:

"On 04-02-99 8:34:21 PM, Frances wrote:

> Any one else feel like their drowning under too much information?

>Yes!!! "

The voting mechanism clearly had some effect. One student wrote:

"I ain't got no votes. Anybody that votes for any of my messages gets
double back from me!

How can anyone resist an offer like that? And
don't worry I'm sure it's well within the rules."

The student did not get any votes this way, although it demonstrates how self-organisation might lead to a structure that achieves a non-educational goal, unless that educational goal is to understand political processes.

It is possible that the voting mechanism was helping to sort out the useful from the less useful, although this was not a controlled study and no firm conclusions should be drawn. Certainly there was a significantly greater depth of threads than for the previous year’s exercise, regularly extending to fifteen levels deep. The ‘Ask Jenny’ thread only attracted six levels of threads as opposed to the previous year’s ten, which suggests either that the replies were more appropriate or that it was easier to identify those which were useful. There is unfortunately no easy way to identify from the results whether the causal factors for the differences in thread depth were due to voting, the interface or the nature/prior learning of the group. A controlled experiment would need to be carried out to test this further.

The maximum number of votes for usefulness given to any message was twelve, with one hundred and thirty five of four hundred and thirty-eight messages receiving at least one vote and sixty three of those receiving two or more.

A large number of rated messages (sixty-nine of one hundred and thirty five) related to the ‘Ask-Jon’ and ‘Ask-Jenny’ threads, but that left a significant remainder of messages (sixty-six) posted by students in the assessed sections and found useful by other students, which provided some indication of the level of trust in the responses as well as their usefulness. This seems to demonstrate some potential for self-organisation.

In total, eighty-seven messages posted by thirty-three students were rated (the remainder of ratings being for my messages), a sizeable proportion of the total. However, there was a strong bias towards messages posted by a small subset of students. Again, this is clear evidence that a small community of trusted experts was emerging from the crowd. The system did not give any clear indication of who these experts were apart from on a per-message basis. If this system were to develop further it would be interesting to explicitly draw attention to well-rated experts, perhaps with a ‘see other messages posted by this user’ button, or with a ‘top posters’ list. Such a system might help to encourage a higher level of competition amongst posters, giving some sort of reward (peer-approval) for success.

Thirty-one of the one hundred and thirty five rated messages were posted at the top level of the hierarchy, and twenty-five of these were questions in the ‘Ask-Jenny’ or ‘Ask-Jon’ themes. Although none of these was highly rated (a maximum of two ratings) it is interesting that questions should be rated at all. It seems that the rating system was being used as a reinforcement mechanism, drawing attention to the significance of the question. This is borne out by four ratings being given to a message in the meta-discussion asking for separate themes for each seeding question (see next paragraph). A system designed merely to make identification of relevant information easier was adapted to further uses, rather like Gould’s spandrels . This is clearly the kind of mechanism that needs to be encouraged when designing evolvable learning environments.

When the discussion started a single assessed theme was divided into threads. This proved unpopular with a number of students who also felt that there was insufficient range in the topics addressed. Based on their feedback, a further five themes were added. This proved popular, and prompted the following message to the meta-discussion from one user, which also provides some evidence of crossing boundaries:

"I think the discussion groups and the resource database are an
excellent method for interactive learning and better understanding of the
networking issues. The discussion group was improved by dividing into
themes. Some of the links in the Resource database have been very useful
for gathering material for my dissertation "Wireless data transmission". "

The fact that I had to play the role of the theme designer was a limitation of the design of this exercise. However, it is easy to see how students themselves could create such themes collaboratively. To make such a system fair and effective, a means of pruning and collaborative identification of good themes would need to be devised, perhaps using a simple rating system, perhaps even through the discussion mechanism itself.

Given that the students were using an experimental system, there were a number of technical problems that were addressed via the meta-discussion. However, there was no evidence that the technology posed any greater difficulties to the users than the previous year’s use of the newsgroup.

2. Some conclusions to the third study

1. Conclusions to this chapter

2. : Planting gardens in the stuff swamp
1. About this chapter

I have already identified some of the features that typify complex learning environments and we have seen how principles of self-organisation might apply in existing networked learning environments. This chapter is concerned with network-based systems that actively assist groups of learners to self-organise. It will largely take the form of a review of a range of tools which, to a greater or lesser extent, enhance the interactions of their users to bring out form or structure. The discussion will include reference to a number of systems which are not intended as learning environments but which none-the-less might assist learners and are to some extent self-organising. Houghton observes that "the qualities and durability of the system depends heavily on the systems of communication and information transfer" . I am seeking ways of bringing about communication and information transfer which will tend towards qualities that will benefit learners.

Although it will sometimes be difficult to identify the degree of self-organisation in a system, I will apply a simple heuristic to provide a coarse level of sorting. A self-organising learning environment will be one that achieves some level of organisation as a result of the interactions of its users, that organisation being unpredictable in advance by designers of the system. It is predictable that a newsgroup will exhibit organisation in the form of hierarchies, but it is unpredictable how those hierarchies will be used and the effects of feedback within the group. If there are further patterns of organisation discernible within the newsgroup, it will therefore be a candidate to be considered somewhat self-organising. To this end, positive and negative feedback loops will be sought where these seem relevant. Lessons learnt from the systems examined in this chapter will feed into the design of the CoFIND system, described in chapters 6 and 7.

2. Structuring of content

1. Communication

At least some duplication of the functionality of a teacher could be accomplished by generating metadata that fit the three general forms characterised above. However, as we have seen, teachers do more than structure resources. They are involved in mediating communication, in responding, answering questions, assessing work and so on. Conversations, especially in groups, may be seen from a systems perspective as individual systems, each of which will be describable in terms of the same kind of metadata that would be used to describe any resource. However, within those conversations are relationships and meanings which themselves constitute a subsystem or set of subsystems, and which are generally far more dynamic than those of other resources. Metadata that relate to communication may be seen in terms of what, why and (especially) how is it related to other communication, but the potential complexity of interactions makes it a special case. This is exactly the distinction which is drawn by Michael Moore with his theory of transactional distance, where structure struggles with dialogue- as one goes up, the other goes down and vice versa. The problems with communication are quite distinct, therefore, from those that relate to the structuring of resources and, in a large networked environment like the Internet, are to do with disentangling the threads and reducing the noise of the parts of the conversations which are irrelevant. Although this may seem superficially similar to the problem of dealing with more static resources it is not. Conversational processes are always situated, with statements, expletives, performative utterances, civilities, replies and replies to replies and so on all contextual and barely meaningful outside that context. I will therefore look at systems that organise communication separately from those which organise other resources. Sometimes the distinctions will blur and there will be many instances of systems to do with the organisation of information, which are also communication systems. Indeed, they would not be included in this thesis if they were not, although often that communication will be artefactual, as abstract as stamping a chad out of a voting slip to communicate a vote for a president. In fact, some might disagree that this is communication at all. Franklin, for example, is quite specific in his use of the word:

Communication, in the sense of the word I intend, requires the sending and receiving of signals. Is the bobcat signaling when he leaves tracks in the snow? I think not. How about the tigress marking the boundary of her territory with urine? I think so. And the difference? Intention, in the folk psychology sense, not the philosophical sense.

As Franklin himself admits, there are some very grey areas. For example, animals’ use of pheromones, flowers’ displays for insects and, for that matter, ant trails all seem to communicate a message where the intention to communicate is not apparent. Accepting that the use of the word communication is not always unproblematic, it will be used here to refer to any occasion when a message (however unintentional) is passed from one person to another or to a group. Other forms of communication such as machine-to-machine communication are not relevant to the themes of this thesis.

Communication via electronic means can be categorised into synchronous or asynchronous modes. Such a distinction recognises that the forms that each mode takes may exhibit very distinct structures. I will therefore be looking at each mode separately.

1. Synchronous communication tools

Many Internet-based synchronous communication environments have been enlisted in the cause of learning, including simple Internet Relay Chat (IRC), instant messaging systems such as ICQ and MSN Messenger, MUDs (Multi-User Dungeons) and MOOs (object oriented MUDs), and video conferencing systems such as NetMeeting and CuSeeMe. Many are interesting in the context of self-organisation inasmuch as they allow individuals to create electronic gathering places which can be more or less attractive depending on the number of visitors to them. Preece observes the self-organising nature of this:

Too few people, and there will not be sufficient discussion to retain people’s interest and draw them back; too many participants, and the community may become chaotic, and people will start to leave.

Like the well-trodden footpath, success breeds more success and a form of stigmergy can take place. However, multi-user synchronous communication mechanisms are noted for their chaotic nature, with chats often tending towards the surreal as users type in their messages in (not quite) real time. One or two systems assist in bringing some sense of order to this chaos, through processes of self-organisation.

1. Example: Odigo

Odigo is a client-side real-time chat application that allows annotations and discussions to centre around Web sites but with the added sophistication that it indicates who is visiting the site now. This allows stigmergic clustering to occur as users congregate on busier sites. It also incorporates a ‘"What’s hot" mechanism to identify topics of current interest to those online, based on popularity, another subsystem which allows emergent behaviour to occur.

2. Example: ChatCircles

A particularly elegant system for real-time chatting is represented by ChatCircles , a real-time chat system akin to IRC (Internet Relay Chat). It provides a graphical interface to a text-based chat system, which represents every user as a coloured circle on the screen. Those who are actively chatting are represented by bigger circles, whilst those who have not contributed fade away to small dots to represent their online presence. Users may move their circles around the screen. They are likely to move to areas where there are larger circles because of stigmergy, a positive feedback loop where success breeds more success. The need to move is encouraged by the fact that users cannot ‘hear’ what is being discussed unless they are nearby, within the ‘zone of hearing’ of a given circle.

Histories of chats may be recorded and played back through another interface which represents individuals as vertical time lines, the messages they send shown as horizontal lines of varying length bifurcating their timelines. This graphical representation shows the shape of particular conversations, hence,

The sequence of growing and shrinking circles creates a pulsating rhythm on the screen that reflects the turn taking of regular conversations.

The system is designed without central control, and yet encourages and graphically depicts the emergence of order within it. Though not explicitly stating the goal of self-organisation, the authors are very keen to avoid anything that smacks of a skyhook,

Color in ChatCircles is purely discriminate. Indeed, it is one of the challenges in this project to find a palette of colors that will not inadvertently suggest meaningful interpretation.

As a tool for mediating conversation in a learning environment without the aid of a moderator or teacher, ChatCircles has many strengths. By encouraging self-organised clustering, particular areas of interest might be formed. However, there is no clue as to the nature of a given discussion. It is as likely to be about visits to the pub as it is to be about a learning topic. As so often happens, this is a system which self-organises but which does not necessarily (nor even probably) do so for the educational benefit of its users. It would none-the-less be interesting to use this in an educational environment as a focussed group of learners might well achieve a form of self-organisation related to learning goals as much as social needs.

2. Asynchronous communication tools

Asynchronous communication tools are certainly the most widely used means of communication on the Internet, inasmuch as they include all varieties of bulletin boards (including Usenet News and web-boards) as well as email. As discussed in the previous chapter, a basic bulletin board offers limited automated support for self-organisation. However, a variety of systems have been developed which, to a greater or lesser extent, allow mechanisms such as stigmergy to provide structures which extend beyond conventional hierarchies.

As previously observed, Turoff et al provide a damning criticism of traditional hierarchically structured discussion fora such as newsgroups, pointing out that there is no support for the semantic links between messages, leading to confusion and obfuscation. They suggest that alternative representations of the relationships of one message with another should be created:

CMC systems with tailored content oriented discourse and visualization structures can become the foundation to support large scale ‘electronic community systems.’ This has been defined as ‘a computer system that encodes the knowledge of a community and provides and environment that supports manipulation of that knowledge.’

Essentially, Turoff et al are suggesting the creation of metadata that represent semantic information about messages and their relationships with other messages. They are explicit about the contextually situated nature of these metadata, requiring that they be generated according to the needs of the group to "evolve discourse structures." The use of the word ‘evolve’ in this context appears to be the weaker more general sense of the word than that of Darwinian evolution, but the idea sounds promising in principle. The production of systems that are actually usable by the thousands of people Turoff et al imagine might be involved in discussions is still some way off. There is a need for intuitive and simple-to-use systems that embody the kind of metadata Turoff et al are asking for.

3. Computer Supported Collaborative Argumentation tools (CSCA) – e.g. D3E and POW!

The concept of Computer Supported Collaborative Argumentation applies to any of a wide range of technologies designed to support the argumentation process. In particular, they offer assistance with a range of problems defined by Rittel as ‘wicked’ . CSCA tools seek to provide mechanisms for structured discussion and argumentation to help organise discussions and avoid the problem of losing the thread identified earlier in the analysis of newsgroups.

1. Example: Journal of Interactive Media in Education (JIME)

JIME uses the D3E engine developed by the Open University. Discourse is centred on a document, immediately providing a level of structure not found in a free-form forum. Messages are posted much like those in a traditional discussion, but with the option to add icons which represent the relationship of a given message to its predecessor, especially in terms of agreement/disagreement (figure 5.1). Thus, a further layer of metadata helps to distinguish the form of an argument, without that form being explicitly imposed from above.



Figure .1 – JIME

A similar principle may be found in a large number of Web-boards, where visual cues (usually icons) give some indication of the content of a message. Icons demonstrate the intended uses of each message, and cross-hierarchy clusters are often apparent at a glance. An example of this may be seen in figure 5.2. Again, these clusters arise as a result of individual postings, not through design.



Figure .2 – a typical web-board

2. Example: POW!

Like D3E, POW! (Perspectives on the Web) is designed as a collaborative argumentation tool . Users of POW access information via ‘perspectives’. Perspectives can be personal or shared, and may exist in hierarchies, including notions of inheritance and multiple inheritance. Personal perspectives are constructed through inheritance from team or class perspectives. A perspective will contain a variety of electronic information sources, such as emails, Web pages, annotations and notes. Individuals add linking structures, but as individuals can browse and incorporate other linking structures from other users within groups, different perspectives may be collaboratively constructed. The potential for such a system is large, but Stahl has commented (in a personal discussion) that the students do not always find it easy to use. The combination of complex metaphors and a rich environment requiring a certain amount of learning makes it less self-organising than we would wish. The issue clearly arising out of this is the need for simple metaphors and easy-to-use, adaptable interfaces.

4. Annotation systems – e.g. uTok, Comentor, CoNote

When Tim Berners-Lee first designed the protocols and tools which now make up the Web it was in some ways a more sophisticated beast than that with which we are now familiar. In particular, it was not a simple publishing medium, as it allowed for groups working together to collaboratively annotate and link items. The most common implementations of the system lost this ability, but several tools have been created to restore such a mechanism. Typically these tools allow a user to share annotations with other users, although they often have other powerful mechanisms to enable group working, such as chat programs and presence indicators (who else is here?), helping to provide a stronger sense of community. This is particularly important in the anonymous world of the Internet, where it is often difficult to gauge the reliability of a particular user. The ability to interact is a step towards building trust.

Annotation systems work on a similar principle to that of D3E, whereby communication occurs within the context of a document, which in these cases can be any web page. Virtually all annotation systems make some use of stigmergy as a means of self-organisation. Visiting an annotated page will give a fairly instant clue to its popularity, based on the number of contributors and quantity of annotations. Like a footpath in a forest, use encourages more use.

As well as providing annotations, uTOK incorporates a voting mechanism of some sophistication, which allows users to set up their own polls (Figure 5.3) on their notes, which can relate to the Web sites they were viewing when the poll was created if they wish. Users seem to use this to discover consensus opinions on everything from interface issues to opinions of movie characters.

The ability to vote provides a particularly powerful mechanism which results in low-rated notes being actually removed from view. This is an evolutionary struggle which organises itself to remove poor messages, thus leading to messages well adapted to the communities that use them. A more basic form of collaborative filtering also applies to the selection of groups in the first place. If there are too many topics of insufficient interest, then a user will leave a group. This works stigmergically, and like nesting birds or termite mounds, the groups will centre themselves to form a self-organised community.



Figure .3 – uTOK

The CoNote system is a shared annotation tool with the specific goal of allowing Web pages to provide a context for discussion. Based on a proxy system that intercepts calls to specific servers and adds appropriate annotations, CoNote allows the structure of the Web document to shape and give structure to the discussion that surrounds it. CritLink Mediator (http://www.crit.org/) operates in a similar fashion to CoNote. Making use of a proxy server it redirects users through its own site to redisplay pages with annotations. A similar application, Third voice (http://www.thirdvoice.com) controversially allows annotations to be added to any Web site, although instead of using a proxy it acts as an add-in to the client’s Web browser. In each system, annotations act stigmergically, like the mud piles of termites: more annotations attract visitors to wonder what all the fuss is about.

The future of annotation systems now seems assured with the development of the Annotea project , a product of the W3 Consortium which makes use of XML standards such as RDF (Resource Description Framework) to store annotations relating to any given document and Xpointer to indicate exactly where in the document the annotation should be placed. Due to its integration with RDF, Annotea systems will be searchable across a distributed network, providing potential for large-scale collaborative efforts, such as virtual distributed peer-reviewed journals of the sort proposed by Edmons . However, as we shall see when discussing collaborative trails, the nature of the Web can make annotations held separately from documents become disconnected and out of date if the document is not edited with great care.

5. Knowledge bases – e.g. the Answer Garden, Knowledge Garden, Mecpol, Q&A sites

A collaborative knowledge base lies somewhere between a discussion forum and a set of static, published web pages. Collaborative knowledge bases grow, hence leading to a proliferation of garden and evolutionary metaphors in their names and descriptions. The original Answer Garden was the PhD work of Mark Ackerman . Like many good ideas, the concept was essentially simple. A database of answers to questions was created by a group of experts. As users of the system asked questions they were given a list of possible answers in return. If none of these answers gave them the help they needed, they pressed an "I’m not happy" button, which redirected the question to an appropriate expert. The expert’s answer was then incorporated into the database, making it more likely that the next time somebody asked a similar question that the answer would be found. In this way the system grew and adapted to the needs of its users. However, this was a garden which never stopped growing, lacking much in the way of negative feedback, leading to a potentially tangled and overwhelming amount of available knowledge. In particular, Ackerman and McDonald draw attention to the need for contextual information to identify the relevance of the answers. This context dependency in learning has already been noted several times and comes back to haunt us in most attempts to provide automation of structured knowledge. Another issue which I have observed in my own experiments and which is raised by Creed is that of the need for anonymity. This is closely linked to the notion of trust and the social nature of any interactions.

Answer Garden 2 sought to address these problems . The main mechanism for organisation is a structured escalation mechanism, allowing messages to be posed to known or trusted sources before being passed on to the next level of help. From a self-organising point of view the elegance of this system is that any user of the system, not just an accredited expert, might give an answer at the first level. Thus, learners are able to take full advantage of teachback, providing a feedback cycle that benefits both the helper and the helped, increasing the knowledge of each. Answer Garden 2 allows for the insertion of rule-based or other systems to find appropriate experts, allowing for the possibility of a collaborative mechanism for identifying such experts, which would help to introduce the cycles of feedback needed for self-organisation. The system also allows for information culling, organising and distilling. These processes are user-controlled, thus affording a more interesting dynamic of collective gardeners. However, a garden is not like a jungle. At its heart is a Stalinist regime that dictates the forms that interaction will take. Although content springs from the users, organisation occurs due to the intelligent algorithms provided by the programmer. It is possible to consider that, with its open structure, it might one day itself be able to evolve.

One notable lack in the original Answer Garden was an effective means of pruning, although this was somewhat improved in Answer Garden 2. This problem is being interestingly addressed by a group of BT projects, which started with a system called the Knowledge Garden (recently renamed to "Information Garden", presumably due to the prior existence of a far less interesting product of the same name) which is currently wrapped around a system called Jasper. Jasper incorporates a VRML interface to an information garden. In the Information Garden, clusters of associated URLs are visualised as flowers growing . Leaves that are little used wither on the stem and eventually fall off. The colour of flowers codes for document status (updated, unchanged, dead) and, when selected, the flower stems wave for a while like real flowers that have been touched. A visitor seeking interesting information is thus drawn by colour and movement created by other users of the system. Each user has an avatar, the position of which records users’ perceived availability based on mouse clicks, keyboard presses and the last time their emails were read. Users can prune the stems that are not wanted and even take cuttings to grow in their own private gardens. An interactive space is created where the environment is formed out of the active participation of its users, a place where knowledge can be shared independently of physical space and formal organisational groupings. As a tool to assist organisational learning this is powerful and could easily be adapted to groups of like-minded learners.

The concept that underlies both answer and knowledge garden is that of an organic yet tamed system. These are not jungles, although they are subject to the rigours of evolutionary selection. However, nothing unexpected is ever going to grow here. The Answer Garden will generate better or worse answers; the Knowledge Garden will provide access to resources in a fairly constrained way. If there were an ecological niche of, say, users seeking documents that were published in a given year, then the system would not evolve a new colour or height of flower. This would still be in the hands of the designer, the head gardener who controls the available variations. In our search for systems that evolve we have taken an interesting step in the right direction, but we are not there yet. The teachers have gone but they have left footprints and strictures in the way of a maximally self-organised learning environment. When the gardeners leave the weeds should grow and strangle the delicate flowers, but this garden is too well managed. It would at least be nice to see something more like a Capability Brown design than a formal arrangement.

1. Information flocking

In a related BT project to the information garden, a metaphor of schools of fish is used to indicate similar clusters . This system is based upon the observation that similar fish tend to swim together. Different coloured fish represent interests of groups of users, and their clustering behaviour indicates similarities between individuals and of different groups (figure 5.4). This harks back to the work of Reynolds, who demonstrated that emergent flocking behaviour can be created by the interactions of three simple rules- collision avoidance, velocity matching and lock centring , demonstrating yet again the complex systems that can be generated by simple rules acting on many inter-related parts. By tuning the weighting of the three rules, different flocking behaviours may be produced, like those of bats, sparrows or fish for example. Although the information flocking experiments are a form of data mining, intended primarily as a way of visualising existing relationships, if they are successful then the results will affect the actions of those whose relationships are being modelled, allowing the emergence of connections where none existed before. This is another example of the operation of stigmergy in the service of self-organisation.



Figure .4 – information flocking

2. JITOL & MECPOL

JITOL (Just In Time Online Learning) was a visionary project designed to assist collaborative learning by allowing people to interact and (more interestingly) to reify their interactions to generate knowledge from debate . A knowledge base served as the reference for interactions and was itself updated by the outcome of those interactions in an iterative, recursive process that was somewhat self-organising as a result, albeit with a certain amount of effort on the part of those involved. The system providing a framework but not very much conceptual support. This was partly remedied in the MECPOL (Models for European Collaboration and Pedagogy in Open Learning) project, which might be characterised as JITOL TNG. At the centre of the MECPOL project is the concept of the "evolving knowledge base" (EKB). The EKB stores information about foreground resource (online resources directly available), background resources (off-line resources) and meta-resources (aids to finding background and foreground resources such as bibliographic databases, information about other JITOL users, on-line reviews etc). The EKB sees everything in the system as a resource, including the interactions between them, for instance an exchange of messages between tutor and student. In the words of its authors:

The key feature of the EKB, then, is that it reflects and embodies the changing beliefs, needs and activities of the community of practice which it supports. It represents a growing, self-improving resource, created through the collaboration of knowledgeable learners and tutors. It stands in sharp contrast to alternative forms of IT-based learning resource, which at their worst are expensive ossifications of tutors’ interpretations of the world.

The authors of MECPOL present a variety of models for the continuing development of open and distance learning (ODL). They warn against uncritical use of the simple prototyping model on the grounds that the needs of each cohort of learners will be different from those of all others. Lessons learnt which would improve the experience for one cohort might be actively harmful to the next. This is a significant insight into the situated nature of learning. Any systems that we allow to evolve should not be considered as universal panaceas. Every cohort of learners represents another ecology.

The notion of making everything on the system a resource, including user interactions, is an interesting one, which would be problematic were it not for the structure which is imposed by the system’s metadata framework. It is interesting to note, however, that the framework itself is fixed and does not evolve. Although we see evolution in the knowledge and how it is presented, the underlying framework cannot be affected by changes within the knowledge itself. It thus breaks Kelly’s law of chunky bottoms and will always dictate the forms that knowledge can take.

3. Q & A sites

There are many question and answer (Q & A) sites on the web, which can be rich watering holes for self-teaching learners.

A particularly effective use of a Q & A site is HelpShare (http://www.helpshare.com). This is marketed as the "Micro-Consulting Marketplace". Questions are posed and answered by experts for a price, which is supplied with the question. This makes excellent use of the power of the web, bringing many and disparate people together to do business- in this case to learn and to provide the source of that learning. HelpShare is particularly aimed at solving IT-related problems, an area where, due to the rapid rate of change, education and training opportunities inevitably lag behind the need for them. It incorporates what are described as Trust Algorithms, to allow questioners to assess the reliability of answerers’ responses and for answerers to assess the questioners’ ability and willingness to pay promptly- a self-organising feedback mechanism. However, it suffers from a problem that besets all such systems, the cold start phenomenon. John Harney writes:

A virtual community like this can only succeed if three things happen- the number of members is sufficiently large, the content is sufficiently rich and the relationships that members form are sufficiently nurtured. Obviously the three phenomena are interdependent- lots of members contribute more content which in turn makes it harder for members to give up the relationships they form with various experts and defect to another community.

Harney’s analysis succinctly explains how this system, though self-organising to a large extent, is not autopoietic (i.e. it cannot pull itself up by its own bootstraps). The system clearly has to reach a point at which it becomes useful but prior to which it is close to useless and may be positively harmful, requiring a much larger investment than the immediate return. This investment is inevitably compounded by the need to provide information to feed the Trust Algorithm. It is just this sort of complexity that continues to make people search for a grand designer to life the universe and everything. Evolution almost certainly managed it in small steps, each of which was self-sustaining . There were none of Dennett’s skyhooks along the way . It is hard to see how such a system could start from scratch in the chaotic jumble of the Internet.

The missing steps needed by HelpShare have been taken by Yahoo (http://www.yahoo.com). Yahoo’s Q & A site does not cost its users anything, although the answers are probably not as valuable as those found on HelpShare (yet). It is organised by Yahoo staff, the only chaotic element being the stream of questions flowing by. It none-the-less shows a route through which the likes of HelpShare could be successful, through a symbiotic existence with something already valuable enough to visit, with a readily available mass of visitors in sufficient numbers to make it self-sustaining. In this sense it might come to resemble Margulis’s account of the role of symbiosis in evolution , joining two already successful entities to make a third, more successful species.

2. Recommender systems
1. Looking for the good stuff- descriptions of utility

Finding useful resources and structuring the students’ approaches to them is central to the process of learning and is one of the key roles of a teacher, be they virtual or otherwise. We have already looked at some ways in which structure can evolve within interactions with the support of computer-based tools as an aid to communication. I now turn to ways in which computer support can provide mechanisms to assist the structuring of content through the interactions of groups of users. It is worth noting that we have not altogether left the world of communication, inasmuch as any collaborative process implies some form of exchange of information.

2. Seals of approval

1. Collaborative filtering

We have already seen how the Web provides us with numerous ways of communication that were simply impossible before the advent of a large, public network. Collaborative filtering is a particularly powerful example of one of these ways.

The concept of the collaborative filter dates from the Tapestry system, a product (like so much of importance in the world of computing) of Xerox PARC in the early 1990s . Tapestry allowed users to add annotations and ratings to messages, thus enabling a community of users to filter messages using a system based on the intelligence of human beings and the raw processing power of computers . This system and its offspring, sometimes termed ‘recommender systems’ or ‘social filters’ come in many different forms but all are based on a simple premise – that the recommendations (explicit or implicit) of others can help us to find what we are looking for. Collaborative filters are organisers of knowledge, with a classification scheme based on the preferences of others. Typically, such systems are used to recommend resources based on a matching of users’ interests, but the visionary Sasha Chislenko noted that they could be used in a much wider context, of particular interest to us being the idea that "educators would pay special attention to things that students find too difficult or boring" .

2. Forms of collaborative filter

There are several ways to classify collaborative filters, based on how ratings are discovered, the forms these take, the algorithms they use. For our purposes, they can broadly be split into those which use implicit ratings (tacitly expressed metadata) and those which use explicit ratings. Between these two extremes there are many partial and hybrid systems.

3. Some examples of collaborative filtering approaches to the discovery of resources and their value

The classic explicit rating system was probably Firefly, largely the brainchild of the late Sasha Chislenko, and prior to its purchase and subsequent abandonment by Microsoft among the most successful recommenders of movies on the Internet. Users entered ratings for a number of movies they had already seen until there were sufficient ratings to be able to match them with those of other users. It was an early instance of an Automated Collaborative Filter (ACF). An ACF explicitly seeks patterns of matching rating behaviours between individuals, not relying upon a simple cumulative vote but instead extracting mined similarities.

Beehive was a tacit rating system designed to discover communities of interest, based on how it recorded the behaviour and interactions of members of communities of practice .

GroupLens is a system designed to assist users to cope with the large amount of information carried in Usenet Newsgroups. Its strengths lie in the hybrid approach it adopts to collaborative filtering, taking explicit ratings, implicit information such as time spent reading articles and content-based information extracted from the messages themselves . The underlying principles of the system have since become transformed into the Net Perceptions product (http://www.netperceptions.com/) widely used commercially.

1. General knowledge retrieval with recommender systems

Recommender systems lend themselves well to identifying personal preferences, but have often been less successful when applied to problems of general knowledge discovery. A sophisticated approach that lends itself well to focussed groups of learners is WebTagger . This is a collaborative bookmarking system designed to address various identified problems with existing bookmarking systems, particularly those of a monolithic or hierarchical structure, lack of facilities for sharing, inability to rank and a clumsy navigational access model. It incorporates a storage system based on a lattice structure, where URLs are stored at different category nodes with a weighting applied to their utility. Users can select one or more categories and rank URLs within them, thereby providing a feedback mechanism that can help refine further searches. The system incorporates a genetic algorithm that purges nodes when their frequency of access drops below a specified level.

WebTagger provides unique views for each retrieval context, the views being generated by the combined ratings of all the system’s users. It is an adaptive system with a powerful set of mechanisms for self-organisation. There are many similarities between this system and CoFIND, the experimental system I shall report on later.

Some of the team responsible for WebTagger went on to produce DIAMS, a Java-based system integrating knowledge-bases, neural networks and genetic algorithms "to develop an ecology of users and agents that evolve over time" . This system is essentially an enhanced version of WebTagger with added rule-based intelligence, which in a sense dilutes the ability of the system to self-organise, instead introducing a designer, albeit an adaptable and flexible one.

4. Recommender systems explicitly designed for learners
1. PHOAKS

The PHOAKS (People Helping One Another Know Stuff) system (http://www.phoaks.com) employs an ingenious form of implicit collaborative filtering. It sorts through news groups looking for URLs. It works on the assumption that the more often a resource is mentioned the more useful it is likely to be, as people seldom refer to URLs of no use. The results of a search on a given term are displayed in order of popularity or reference. This idea has been taken up with a vengeance by the extraordinarily powerful Google (http://www.google.com) search engine, which rates the popularity of a given resource by the number of times it has been referred to in other Web pages. Such systems embody self-organising principles by combining the actions of individuals to create a pattern not envisaged by those individuals. CiteSeer (http://www.citeseer.com) works in a similar fashion to PHOAKS in that it performs a citation analysis of academic papers.

2. LON-CAPA

The LearningOnline network is an attempt to create a market-driven infrastructure for resource sharing in learning systems. This incorporates a large number of mechanisms with a greater or lesser amount of centralised control, from course management tools to communication and collaboration mechanisms. Its most interesting component from our perspective is the use of genetic algorithms in the provision of the student interface to the system. A form of adaptive hypertext, the system will learn from learners’ previous selections to automatically customise courses to their needs. The authors hypothesise that it should be possible to cross-compare different students’ preferences and learning outcomes.

By offering the resources that were deemed ‘better’ to the students and encouraging the instructors to supersede the ineffective resources, the human-computer interaction could lead to more effective virtual university classes. In system science, this type of creation of structure via selection rules and feedback loops is known as ‘self-organized criticality.’

The explicit attempt to produce a system which self-organises is interesting and reflects many of my concerns that I attempt to address with the CoFIND system. In particular, this is not a simple feedback loop, inasmuch as the control mechanism is itself able to adapt. Good resources result in successful selection; bad resources are forced to adapt. The effect of this adaptation has the potential to change the way in which resources are rated, because the learners have learnt as a result of the rating process. It is easy to see how such a system could approach self-organised criticality, although it would be interesting to see how successful this system turns out to be. For the feedback to be effective it requires active participation on the part of the assessors and the producers, something which is far from guaranteed. In addition, the mechanism of feedback gives little indication of the sources of the learners’ displeasure. Knowing that something is bad is part of the way there, but knowing why it is considered bad is probably more important. This system does not provide sufficient meta-information about itself, so the self-organisation is only going to occur along a linear plane, with a restricted number of connections between nodes.

5. Capturing knowledge

Tacit knowledge is notably tricky to capture, as by Polanyi’s definition it is inexpressible. The notion of inexpressible knowledge is perhaps contentious , but it is certainly true that some knowledge is extremely hard to capture and express. Stenmark has built a system which seeks to externalise tacit knowledge by capturing patterns of behaviour, the documents we read, the Web sites we visit, the people we contact and so on. His system makes use of agents that capture such information into profiles and allow it to be shared amongst other users of the system. Hence, it is possible to discover users across an organisation with similar profiles and matching interests, not through any explicitly stated similarities but through matching tacit knowledge. As with most recommender systems, the system suffers from a cold start problem, with the need for agents to be trained before useful results could be returned. Despite this, such an approach is promising, inasmuch as it captures behaviours that are not explicitly stated. In terms of complex adaptation however the system is just a means of automating the process of bringing people together and does not exhibit much particularly emergent behaviour per se, although it is easy to see how it could generate clusters based on the habits of its users.

6. Problems with recommender systems

Collaborative filters and other recommender systems provide excellent candidate methods for replacing the role of the teacher in a self-organising learning environment, but a number of problems present themselves.

1. The cold-start phenomenon

1. Descriptions of relationships

The relationships between different objects of study have already been identified as an important contribution made by teachers to the educational process. Such relationships may relate to sequence, as in what should come next. This is always going to be a complex problem and one that requires expertise in both subject matter and pedagogy. It is not always clear what that entails. There are usually many ways to organise a particular piece of learning material. Should it be based on the conceptual mappings of the subject expert, a simple to complex pattern to assist the learner or any one of a vast range of other alternatives? In many ways the answers to these questions are what distinguishes good teaching from bad, and it is hard to conceive of such order emerging out of mindless interactions. The key point goes back to a simple maxim stated by Ramsden:

An effective course will have its material arranged in such a way that the issues generate confidence and interest in students.

There is no simple prescription for this, and the needs, skills, learning preferences or learning styles of the learner will play a large part in determining the optimum course through a given body of material. I have already observed that the Web provides a mass of learning resources from which to choose. Many of these resources are already shaped according to their authors’ beliefs in their utility. Learners are not faced with a blank palette to paint trails. Instead, there is a rich diversity of possibilities from which to choose. The problem is not in creating the trails, but in choosing the right one at the right time. Like so many of these problems, there is an issue of scale. Within a given course there may be a range of topics and within each topic a set of subtopics and for each subtopic an optimal way of addressing the subject matter. If, as Vygotsky tells us, developmental processes lag behind learning processes (the gap between them being the zone of proximal development) then it would be foolish to learn, say, differential calculus before basic arithmetic. Within the topic of basic arithmetic, there may be a set of sub-processes leading from addition and subtraction through to multiplication and division. Within each of those processes there may be steps from simple digital manipulations through to complex fractional calculations. Identifying topics and subtopics is more complex, therefore, than simply selecting the best path. Equally, the effects of different learning styles (perhaps I should use a more neutral term such as ‘learning preferences’) might make the best path for one student very different from that which is best for another. Paths branch and intertwine in myriad ways. It is important to consider how such paths might be constructed emergently.

1. Pathfinding and trails

Look at every path closely and deliberately. Try it as many times as you think necessary. Then ask yourself, and yourself alone, one question...Does this path have a heart? If it does, the path is good; if it doesn't it is of no use.

1. Bush and the Memex

Vannevar Bush is widely hailed as a visionary, the creator of the concept of hypermedia in the form of his imagined machine, the Memex . The Memex was beyond the possibilities afforded by the technologies of the time, but conceptually incorporated many of the features that we now recognise in hypermedia’s most prevalent form, the World Wide Web. One important aspect of the Memex which tends to be missing, however, is the simple ability to record and annotate paths ("associative trails") through linked media. This is not to say that such a thing is impossible, merely that the process of authoring for the Web is not and intuitive nor a particularly widespread skill and even amongst practitioners is usually done alone, if not in private.

1. Collective Mental Maps

Francis Heylighen has been researching in areas which closely overlap with the concerns of this thesis, seeking emergent properties of the actions of groups of individuals to create what he describes as Collective Mental Maps (CMMs). His aim is to solve the typical problems which occur when a group of intelligent individuals attempt to combine their knowledge as, for example, we find in a committee. The typical committee exhibits no more (and often a lot less) intelligence than its individual members. Problems with such systems are manifold, ranging from pecking orders, to differences in mental maps, vocabularies and conceptualisation, to difficulties with volumes of data. A typical solution to this problem is an organisational hierarchy, of which Heylighen writes:

Although no president, chief executive or general can oversee or control all tasks performed by different individuals in a complex organization, one might still suspect that the intelligence of the organization is somehow merely a reflection or extension of the intelligence of its hierarchical head.

Heylighen’s solution to hierarchies’ weaknesses is to look for the computerised equivalent of the archetypal stigmergy of ant trails. In essence, a weighting is applied to sites which have been visited so that those sequences which are used more often are given greater weight, following the application of collaborative filtering algorithms. The main problem with this approach is that it fails to capture the true nature of trails as they apply to learning situations. The first difficulty with Heylighen’s system is pragmatic. It is difficult outside an experimental setting to capture the true sequence of a user’s explorations. His first experimental solution is based on observing user behaviour given a small subset of 150 words. When transferred to the Web, Heylighen’s solution relies on associated documents, not those which follow a particular order of access. Secondly, a learning path is not a continuous sequence. Learning occurs at different rates, with varying amounts of time required for assimilation of ideas, concepts and knowledge in general. A given path may be separated by minutes, hours, days, weeks, months or years. A collaborative path based on movements from one document to the next will not capture this, and so may be of little pedagogical benefit.

2. Adaptive and Intelligent Educational Systems (AIES)and Adaptive hypermedia

1. Assessment

• openness in the educational process;
• self-determination in learning;
• a real purpose in the co-operative learning process;
• a supportive learning environment;
• collaborative assessment of learning;
• assessment and evaluation of the ongoing learning process.

1. Conclusions to this chapter

2. : Burying Teachers in the Stuff Swamp
1. About this chapter

There are inherent features that make resource discovery problematic, especially on a distributed system such as the Web. This chapter starts by discussing some of the issues that relate to discovering resources from the Internet that can often be reduced to issues of metadata in general, in particular the problems of categorisation and classification. This feeds into an explanation of the motivation for CoFIND (Collaborative Filter in N Dimensions). The motivation for the CoFIND system and some of the theoretical reasoning behind it is explained, including methods of self-organisation and its relationship to the previously identified teacher roles. This will lead on to the next chapter, which describes some of the experiments I have performed which have driven the development of the system, as well as a discussion of the omissions and weaknesses which, if overcome, could lead to a truly self-organising learning environment.

2. Problems of Categorisation
1. Metadata standards

Should we all agree to describe our resources to an agreed standard then we might be able to use search engines or other automated tools to find the kinds of things we are looking for. Murray writes:

We are interested in situations where there is an instructional goal....and the resource for achieving that goal is not explicitly known. In a sense, we would like information (called meta-content or metadata...) attached to tutorials which describes what and how they teach so that a search engine can locate an appropriate tutorial on the Web.

Murray believes that such metadata can be specified and codified in a consistent manner. However, there are problems with such a model, as Murray himself recognises when he identifies term ambiguity as a thorny problem. His solutions are pragmatic and interesting, to implement topics (the metadata he is interested in) within controlled or local contexts, to use metadata standards such as the IMS (Instructional Management Systems) and Dublin Core schemas, and to semi-automate rather than fully automate the system. Murray’s solutions skirt around the problem but do not attack it at its core, probably because the problem is essentially irresolvable. I will look at this rather bold assertion in a little more detail.

2. Dynamism and stability

Solomon states:

Even though we often act as if the world were static in the information field, it is in motion and without adaptive mechanisms our knowledge organization schemes are likely to become less and less appropriate over time....knowledge organization schemes need to be seen as living things.

Traditional classification schemas are not as fluid as the world that they model. Classification schema such as the Dewey (DDC), Library of Congress (LCC) and UDC (Universal Decimal Classification) are extremely popular with librarians, forming as they do a theoretically based and somewhat flexible categorisation schema. Developments such as Ranganathan’s facets allow structured classification to be applied to the subtlest nuances of meaning. DDC, LCC and UDC constantly develop but each requires a substantial intellectual investment to be used effectively and modifications and extensions to the schema are the stuff of academic conferences, not arising from the day-to-day usage of them. Pejtersen and Albrechtsen suggest a more flexible view of categorisation as arising within information ecologies and themselves being ecological in nature . Their system is based around empirical analysis of invariant structures within a given information ecology at different levels, which may then be used as a framework on which to hang the more dynamic categories. There are similarities between this understanding and the point noted in chapter three that ecosystems are shaped most strongly by their slower moving parts. Invariant structures are by their nature slow moving. They only appear invariant within a given context; as Pejtersen & Albrechtsen observe, principles of universality have to give way to principles of use and context dependency. Even dynamic and system-based approaches like this are rooted in the world of the knowledge professional, however, and do not necessarily relate closely to the usage of classifications within a given context.

3. Context dependency

1. Information Retrieval issues

With the advent of the web, traditional methods of finding resources using classification and cataloguing to aid information retrieval (IR) have either been forgotten or discarded. Consequently we see a wide range of classification schema, many of which are clearly not produced by knowledge professionals. Yahoo, for example, uses top level categories that are disorganised and barely meaningful:

The librarians at Yahoo! Divide the world of knowledge into fourteen broad classes of unequal status (Education, stands next to Entertainment) and variable types (Arts & humanities, Reference, Regional).

An expansion to the second level in the organization of the discipline Education shows more of the same absence of distinction between various categories of resources, and lack of direction as to what aspect of a complex subject...has been given priority in the classification of the resources.

Faced with such disorder, free-text searching has become the order of the day, despite its known shortcomings . Search engines are growing in sophistication from the simple content-based systems such as Alta Vista (http://www.altavista.com) to second generation systems such as Google (http://www.google.com) and Clever (http://www.almaden.ibm.com/cs/k53/clever.html), yet all are floored by the wealth of information hiding in a mass of dross.

Google is amongst the best and most relevant of all current search engines due to its second-generation collaborative filtering engine and vast collection of resources. A search using Google on August 7^h 2000 for "Ethernet" revealed 1,780,000 results. Narrowing the field by trying "Ethernet tutorial" reduced the number of results to 39,700. This is still an unfeasibly large number of results to wade through. Although many of the results were clearly not relevant, there was a sufficiently large number to make the task of sorting through them fiendish if not impossible. Narrowing the search helped to bring some useful sites to the top, but it is impossible to tell how many useful sites were lost in the farther reaches of the search results. Random browsing through many pages still revealed potentially useful resources up to the point at which Google automatically cut off, 794 results into the search. Adding the word "beginners" reduced the response to 3,200. Unfortunately there is no clue as to what else might have been lost along the way by narrowing down the search query, leading to the loss of "Ethernet 101", "Ethernet Getting-Started" and so on. Also, there are many aspects of Ethernet that we might wish to find out about- there is a big difference between a training manual on how to implement the technology as opposed to a theoretical model explaining the concepts behind how it works. There is a strong case for teaching the skills of searching, but even the best professionals may fail to find the most relevant and useful resources. Part of this problem relates to being able to phrase the right question. When seeking information in a new problem domain, the learner may have insufficient knowledge to enter relevant keywords, a point noted by Heylighen: "in order to select relevant keywords, the user already needs to have a clear idea of how a potential solution would be formulated " . The need for effective IR systems on the Web is a pressing concern.

1. Quality of Information

1. Novelty

1. The Motivation for CoFIND

Dynamic systems of instruction and learning, in information-rich environments, cannot and should not be concerned with the production and dissemination of instructional materials, since there is already an abundance of content around us.

1. First goals

Early thinking about the design of CoFIND focussed on placing learning resources into an ecology that forced them to compete with each other, with the intention that the resources themselves should adapt to this evolutionary landscape. At this early stage I had little idea of the range of existing collaborative filters nor was I concentrating on the roles of teachers. I had observed, like Saba , that there are many good resources for learners on the Internet. As a teacher of networking technologies my subject area is better catered for than many, but the exponential growth of the Internet means that the problems of QOI which I face when searching for something useful will need to be addressed even by those dealing with the most obscure fields of study.

1. CoFIND as it is now

CoFIND is based on identifying patterns of abstracted similarity between resources, not only what resources contain to make them useful, but how those resources are useful, in what context, to whom. The main engine of every version of CoFIND since its first incarnation seeks user-perceived metadata about resources that describe not only positive/negative feelings towards those resources but which characterise what it is about those resources that induces these feelings. Such metadata are referred to in all the CoFIND systems as qualities.

1. Qualities

Qualities are metadata about a resource which describe its value, which would include but not be limited to the kinds of things identified by Valovic as applying to QOI or the selection criteria of Hofman & Worsfield . Typically adjectival in nature, they indicate what it is that a learner likes or dislikes about a resource. Examples might include good for beginners, amusing, helpful or comprehensive. In the CoFIND systems, learners themselves are the creators of these qualities. In keeping with the philosophy of self-organisation, there are no initial constraints on what may be entered as a quality. If a learner decides that what is valuable about a particular resource is the amount of yellow on a Web page then there is no restriction that prevents that quality being added. This can in principle lead to the generation of many qualities of little or no utility to the bulk of users of the system. To introduce constraints, these qualities are allowed to starve to death, out-competed by other, more successful qualities, in a process more than a little akin to that of evolution. Competition is based on two controlling factors, votes and usage.

2. Votes

Qualities are used to vote for resources. For example, users may vote for whether a given resource is more or less good for beginners, amusing, helpful, comprehensive or yellow. If a quality is used to vote for a resource, the quality ‘grows.’ Different versions of CoFIND have used different mechanisms and metrics to define this growth and how it is manifested to the user. Conversely, little-used qualities ‘shrink’ and die of vote starvation.

3. Usage

Qualities are used to provide the sort order for resources. Those qualities which are selected for that purpose ‘grow’. Thus, a weighting is given to yellow if I decide to look for resources which have been rated as yellow, whether or not I choose to rate any as yellow myself.

4. N dimensional filtering

Qualities are the n dimensions in n dimensional collaborative filtering. The need for them is based on the recognition that, in the absence of a trusted teacher, it is not enough for learners to be told that a resource is useful or not useful, good or bad. A learner needs to know how it is useful, in what ways. Learning is about change, with dynamically changing needs arising as a result of learning. Saba says "instructional and learning systems must be able to assess prior knowledge of each learner at the onset of an instructional session and be able to provide differential learning experiences for each student based on his or her prior knowledge of the field" . Saba is correct, but could go further. It is not just prior knowledge but learning styles, requirements, future intentions, personal taste and an indeterminate number of qualities of resources that will influence whether or not a learner will find them useful. The layer of metadata created by qualities allows for a flexible set of behaviours that enable the system to finely adapt to its learners and the resources that are fed into it. Traditional automated collaborative filters are based on matching similar preferences in users. CoFIND makes those preferences explicit, effectively reifying user models so that a user is not pigeonholed into a particular niche, but may select a model which fits his or her current needs, changing that model as learning occurs and those needs develop.

5. Evolution

To attain knowledge, add things every day. To attain wisdom, remove things every day.

(Lao-Tsu)

Qualities play a central role in CoFIND, providing multiple perspectives on a group of resources. Resources rated according to selected qualities are ranked according to how they have been rated. Having decided that evolutionary mechanisms could provide a route into the self-organisation needed, I began to look at evolutionary principles in natural systems.

Evolution requires reproduction with variation and competition for scarce resources to occur. It is fairly easy to see how competition could force out unpopular qualities, but their method of reproduction probably needs closer examination. Qualities are produced by learners, who also provide their only source of food (votes and usage). Once produced, a given quality may be more or less successful. Reproduction may be seen in terms of the number of resources rated. Reproduction with variation may be seen as a process of refinement. If a particular quality does not quite capture a user’s needs it may suggest one that is more appropriate, in which case the user may enter another quality that is more appropriate. This may, or may not, result in more useful resource ranking. The most recent iteration of CoFIND allows for the inheritance of votes from an existing quality when specifying a new one. A user may create a new quality such as "humorous" which can inherit all the ratings applied for the quality "funny." As there are no limitations on choice of qualities from which to inherit, the user may equally select "dull" as the parent quality. Such inappropriate choices will usually lead to useless results (a user seeking funny resources would instead find those which were dull) and thus the processes of natural selection will remove the newly added quality quickly.

The principle identified earlier that evolution tends to occur more rapidly in parcellated populations leads to a useful mechanism for reducing the bloated excesses of results returned by large search engines and directories. A small group of users may make use of such large search mechanisms to find resources, which may then be added to a small database. If a small group with a common learning interest were to collaboratively compile such resources then there would be a much higher chance of relevance than for each individual to wade through the mass of results from a large search engine. Using an evolutionary model, a context-dependent taxonomy might be developed which captures the usage of the group’s tacitly negotiated and agreed set of classifications. In such a system, ambiguities and disagreements between different agents would be ironed out by a process of pointing to examples of instances which satisfy the criteria of the metadata. We may not all agree with the classification of a particular Web page as ‘pretty’ but a majority verdict within the group would allow others to understand the schema being employed by a process of example.

Due to the high degree of relevance to the group, such a system might even be superior to the wide variety of subject-specific directories that are available on the web. This is the approach taken by CoFIND.

6. Stigmergy

The principle of competing qualities means that CoFIND takes advantage of stigmergy, creating positive feedback loops that drive speciation. If a resource appears at the top of a list, it is more likely to be selected. Similarly, if a quality is made more prominent (in CoFIND this is through the use of different font sizes) it is more likely to be used. If it is selected more frequently, it will also be used to rate resources more frequently. Use encourages more use. Instead of an averaging process, positive feedback drives the successful to greater success, balanced by selection pressure and competition.

As CoFIND developed, it became clear that this process might be utilised in a number of different contexts, especially the creation of appropriate paths through a range of subject matter.

7. Topics

1. CoFIND as a replacement for the teacher

1. A theory of instruction should specify the experiences which implant a predisposition toward learning

1. A theory of instruction should specify ways a body of instruction should be structured for easy assimilation by the learner

CoFIND identifies the contents of a body of instruction, resources that are of use, in a manner appropriate to the learners who wish to use them. Implicitly, it specifies the limits of a subject through the use of topics, which it uses to guide a learner through a subject area. Through the mechanism of ratings and qualities it acts as a filter and a recommender of appropriate resources with an efficacy not far removed from, and potentially better than, that of a teacher.

Except in the coarsest of ways, CoFIND does not help to identify ways to structure a body of instruction, nor does it strictly speaking need to. The Web is awash with suitable resources that are structured to assist learners to learn. CoFIND’s strength is that it allows learners to discover those structures that are most suited to their needs at any stage, rather than relying on the skills of a single teacher. However, semantic links between topics and the ability to generate different hierarchical levels would be of some benefit in allowing the self-organisation of structure to occur.

Future developments of CoFIND need to improve the chat and forum mechanisms, so that emergent behaviour can arise of the sort seen in ChatCircles and D3E, discussed in a previous chapter.

2. A theory of instruction should specify the most effective sequences to present materials to be learned
1. Curriculum design

Sequencing in a traditional taught environment occurs on several scales, from the broader topics to the small movements within them. As CoFIND is designed to incorporate resources which will usually have some micro-levels of sequencing (if only in the form of a beginning, a middle and an end) this is not a fatal flaw. To provide ultimate flexibility and to incorporate learning objects, I would wish to see sequences occurring at different levels of a structural hierarchy. The question arises as to whether the mechanisms should be the same or whether there is a difference in kind between micro and macro levels. The current iteration of CoFIND uses a method of real-time sequencing based on dynamic usage, a kind of collaborative trail-blazing, but there is no means of breaking topics down, of identifying sub-topics and semantic links between them. As already mentioned, we might make use of a trail mechanism to show who chose what next, but have already identified that this is an imperfect mechanism for learners whose paths may be weak and confused as they are only learning. At least CoFIND does offer a means for comparing paths created using other tools such as Walden’s Paths .

3. A theory of instruction should specify the nature and pacing of rewards and punishments in the service of learning and teaching.

Rewards and punishments can take many forms, although formative assessment is probably the most prevalent way of providing this. In CoFIND, the same mechanisms that work for evaluation can also be applied to assessment. The rating mechanisms can be closely coupled to assessment criteria in the form of qualities. This allows for collaborative decisions to be made about goals and achievements. The comment system, the discussion mechanism and even chat are all effective ways for feedback to flow through the system in a more conventional manner. Coupled with the ease with which online resources may be developed and posted to the CoFIND system, it has the potential to make an excellent formative (if perhaps not summative) assessment system. Despite the potential, it still requires participants to make active choices about forms of assessment, times of assessment and so on. Although these activities can be discovered in pre-existing resources and the value of those resources can be evaluated using CoFIND, the decision to assess at a particular juncture cannot be supported emergently. Someone, or something, has to make the decisions.

CoFIND was not originally created with assessment in mind. The use of qualities in more than one context, both as an evaluation mechanism and as a means of assessment can be seen as a kind of exaption, with new and possibly unexpected uses developing from a single beginning. There are analogies for this in biological evolution, such as the development of fins in aquatic mammals from limbs, the peacock’s tail for attracting a mate, the eye from simple light sensors and so on. Provision of an adaptable mechanism is a prerequisite for evolution of any sort to occur. The fact that a single mechanism has more than one use within the system is a strength from the point of view of self-organisation, allowing unexpected forms to occur. Unfortunately, although users can base new qualities on old, CoFIND’s qualities do not have the capability to grow in complexity. A quality is a fixed entity. Early CoFIND iterations allowed users to select multiple qualities, but this led to too many connections and too little competition between qualities However, in principle, repeatedly used combinations of qualities could be seen as new structures akin to multi-cellular organisms. In the early CoFIND, there were difficulties in deciding how to treat resources which were rated low for one selected quality yet highly for another, as (for instance) the average of amusing and reliable is not a quantifiable nor a meaningful variable. One solution might be to allow selection of more than one quality with user-defined weightings for their utility (for instance, making amusing a more important quality than reliable when returning results). This might make CoFIND a subtler tool, especially for assessment, allowing users to collaboratively generate not only the criteria for assessment but also a measure of the importance of those criteria. This has not been implemented, as this would increase the cognitive load for the user, as well as the time taken to find resources.

5. Subject expertise

CoFIND excels in the area of subject expertise, by bringing the filtering and generative power of a group of learners to bear on the vast pool of expertise that is afforded by the World Wide Web. In this way it can far outstrip any single teacher in finding suitable resources from which to learn, and can exceed the efficacy of subject gateways due to the selection of resources which are relevant to the group’s specific needs. This has been observable in all the experiments so far and at every level. Saba would approve. He writes:

In the current information-rich environment, the main task in instructional systems design is not choosing, chunking, sequencing, producing and presenting content for a hypothetical mass audience, but helping the learner to find content that is relevant to his or her level of knowledge and understanding of a particular topic within a discipline.

6. Community

CoFIND has built in means of communication. Not only does it have a discussion forum and chat, but the ability to create and share documents, the ability to add comments, even the voting process and the act of participation all result in the flow of information in a communicative process. It is impossible to even use the system without in some sense communicating with other users in the community, if only by contributing to the stigmergic trail. However, the ability to chat and to discuss topics and resources does not automatically create a community. It is not a case of ‘if you build it they will come.’ Communities take rather more to build. We must rely upon the users themselves to build this community, so it is essential to create an infrastructure that makes it easy for them to do so. Like Brand , Preece notes:

The architecture of a house is intimately related to the way its inhabitants live. Cooking, serving meals to a family, or supervising children can be pleasurable or difficult, depending on the facilities available. Similarly, the functionality provided by software and its ease of use- that is, its usability...strongly affects users’ lives online.

One way to go about this is to increase the channels of interaction. CoFIND’s very simple chat room is one attempt to do this, but it is a poor implementation. I have also attempted to implement a system which indicates online presence, but the stateless nature of the HTTP protocol means that it is impossible to know for sure whether someone is logged in, so it can be a bit discouraging to find listed users are unavailable, and the chat system is clunky at best. It would be a lot better to incorporate a mechanism akin to that of ChatCircles. This would not only allow for self-organisation of messages to occur, but might also be linked to specific topics, perhaps even specific qualities within topics, so that focussed discussion could take place on a given area. Similarly, the sort of collaborative touring found in systems such as Nestor might be used to great effect and could even help to provide better paths through resources.

7. Reflective practice

1. Conclusions to this chapter

2. : CoFIND developments and findings
1. About this chapter

The previous chapter introduced the CoFIND system and explained the reasoning behind it and some of its key conceptual underpinnings. This chapter will explain some of CoFIND’s functions in detail and the research findings which drove its development. I then go on to discuss some of the major shortfalls of the system and look at some of the ways in which these may be overcome in future generations of the system.

The CoFIND series of applications represent various attempts to replace some of the roles of teachers through the use of collaboratively generated metadata. This is a description of the process that has led to its current form. This process has produced a piece of software that is different in operation and purpose to any other software that I am aware of. The process of development has played an important role in shaping my thinking and perhaps raises some issues that might otherwise go unsaid.

1. Generations of CoFIND

Table 7.1 provides an overview of the main iterations of the CoFIND system.

Table .1 – the main iterations of CoFIND

Version	Main distinguishing features
CoFIND 1	A simple shared bookmark system making use of qualities, several selectable at once, with voting using a simple "I agree" button
CoFIND 2	Introduced topics. Only one quality is selectable at a time, with a voting system based on a Likert scale. Also contained a discussion mechanism and a means to add non-Web resources. This system went through various minor iterations to correct problems with the interface and to balance the degree of stigmergy it employed
CoFIND 2.5	Topic labels dynamically sized based on a stigmergic/evolutionary mechanism
CoFIND 2.6	Quality labels dynamically sized in a similar manner to the topic mechanism

 

2. Methodology

Appropriately enough, CoFIND has developed in an evolutionary manner, a set of trials and errors, some planned, and some exaptions based on fortuitous spandrels. It started at an early stage in my research from a slim theoretical base which was initially only aimed at looking at evolutionary processes, and has incorporated further theoretical inputs (especially methods of positive feedback based on stigmergy) as I have learnt about them. Many lessons were not to do with issues of self-organisation at all but had to do with HCI (human-computer interaction) problems. However, these are not the primary focus of this research and are not pursued at any great length.

2. CoFIND 1

This instance of CoFIND was originally reported on in Dron et al and in an extended form in Dron et al , which forms the basis of this study.

1. The problem

The problem that CoFIND 1 sought to address was simply to provide a means to organise resources through the combined actions of its users, centring around qualities as the means of speciating different learning needs.

2. CoFIND functionality

1. CoFIND 1 case study

I have already discussed the use of a web-based discussion forum for an exercise performed with a group of forty-two MSc information systems students (see chapter 4). This forum was part of a larger system designed to support the learning needs of the group faced with a problem-based assignment, with CoFIND providing its central knowledge base to assist with learning new facts.



Figure .2 – CoFIND 1 resource display screen

Around half way through their conversion masters degree, the students were required to design a network, a task that would take significant research to complete, with much of the knowledge required being new. The students had participated in a pilot trial of the system for one week prior to the start of the assignment. CoFIND was provided as a database of resources relating to the assignment, and students were offered up to five percent of their marks for their contributions to it, whether adding resources or commenting on or voting for existing resources. They were also asked to use the system to submit the URLs of their finished reports.

I had primed CoFIND with three relevant resources and seven qualities (‘interesting’, ‘useful’, ‘reliable’, ‘informative’, ‘a good gateway to further resources’, ‘of broad coverage’, ‘accessible’) that had arisen in the earlier pilot trial. Throughout the experiment, conscious of the need to avoid behaving as a sky-hook I did not make any further contributions. The starting state of the database is simply part of the shape of the evolutionary landscape. I was looking for the dynamics of the system, what happens next rather than how it came to be.

2. Some Results

The system was very popular and received many positive comments such as ‘we should have had something like this for the entire course’. The main appeal to students was that it was a searchable collaborative resource database- a shared bookmark space for storing useful URLs and associated comments. Thirty-six out of the forty-two students discovered over seventy relevant resources over the two week assignment with a range and quality surpassing my (presumably expert) attempts spanning several years to build a similar list of links. All the students who did not contribute resources made comments on existing resources, indicating that they had spent some time looking at them. Given the allocation of marks to this process, this was not too surprising.

Although CoFIND 1 was designed only to handle resources that can be specified with a URL, it proved flexible enough for one of the students to use it to recommend a book- the ensuing votes and comments showed that this was a very successful choice. As a result of this, later iterations of CoFIND attempted to make use of this exaption by providing explicit support for non-Web-based resources.

Growth of qualities was generally restrained. Only a few new qualities were added to the original seeding values until the final hand-in day, at which point a Cambric explosion of qualities occurred, mainly along the lines of ‘"Brilliant!", "Top !" and "Spot on", all applied to the students’ own work. Until that point I had suspected that the students had not grasped the mechanics of adding qualities and that I had made the process too obscure, but this showed otherwise. With sufficient motivation, qualities were rapidly added. As Wilson and Ryder say, "There must be a legitimate self-interest to sustain individual participation" . Finding the key to that self-interest remains a problem for use of the system. The extrinsic motivation applied here works to a point, indicating that CoFIND may be usable within a traditionally taught environment. It is interesting to note the similarity of the terms used in this last flurry of CoFIND activity, suggesting that the mechanism of reproduction with variation may be valid.

Early questioning suggested that students might not have added new qualities because the existing ones mostly suited their needs and they were avoiding unnecessary effort. Likewise, voting was not as common as I had hoped for. The process was apparently too tedious and went virtually unrewarded. Further questioning has since revealed that this was partly the case, but there was also widespread misunderstanding of how the system was designed to work. Many students appeared to be using the system as though it were a standard search engine, a problem that affected Stenmark’s trials of a similarly tangential system . Indeed, some students added qualities that forced it to behave as a search engine; for instance ‘about firewalls’. Those who added qualities appear to have been those who understood the principles having read the accompanying help and listened to the introduction to the tool. Although I adjusted the interface following feedback from the pilot trial, difficulties understanding the mechanics of voting may still have influenced the results. A similar issue was that the students were at first unsure of the difference between quality selection and inputting search terms. One student made the comment that a resource was ‘good for beginners’ without voting for the quality ‘good for beginners’. Again, there were clearly some user interface issues to be resolved. There may be other reasons why methods of interaction are not understood, most worrying of which is the possibility that users understood the interface but did not want to use it.

Some qualities were a lot more successful than others, with the most popular qualities gaining nearly three times as many votes as the least popular, when measured over an equal period. Qualities that started successfully remained successful, which was an encouraging indication of stigmergy, although there was a significant amount of fluctuation from one day to the next.

1. Useful Qualities

Qualities rose up the list relative to other qualities. For example, the quality ‘brilliant!’ was initially created as a quality of one group of students’ report, one day from the end of the assignment. It was immediately taken up by all the other groups as they handed their work in on the final day (clearly a successful quality when applied to students’ work), and within a day of entry was being applied to other resources and appearing half way up the list.

2. Competing Qualities

There was no clear evidence of the effect of votes for one quality affecting votes for any other, although the fact that students could only vote using a limited number of qualities in a given space of time suggests a priori that competition must have played a role in selection pressure. However, the relatively short duration of this experiment and relatively small number of qualities only allows us to represent the coarsest of trends.

3. Combinations of qualities

To provide flexibility users were allowed to select multiple qualities, with the results displayed being ordered by a simple aggregate of ratings. This could lead to some curious results. For example, if a user selected "good for beginners" and "comprehensive" then highly rated resources for either or both would be returned. This could easily lead to contradictory results. We could imagine that a resource base containing a beginner’s guide to networking and the complete set of technical RFCs for TCP/IP might both be returned near the top of the list. This is an undesirable state of affairs. A more sensible algorithm would have displayed resources which matched both qualities prior to those which matched either, but even then the problem would recur further down the list. Multiple quality selection might be seen as desirable in terms of serendipitous diversity, but it presents huge problems if we wish to return sensible results.

4. Other Problems

Users did not have the ability to change details after a resource had been posted, mainly because I considered that it would be too easy to cheat- for instance, users could amass votes for a known popular site and then redirect it to their own home pages. The flip side of this is that it made it very difficult to deal with resources whose URLs changed. The world of the Web is often ephemeral and fleeting. Resources which were useful today may have changed or vanished by tomorrow. This is especially true in the context of computer networks. Looking back on the resource which were generated by the students over two years ago, up to half of them are no longer valid links and it is impossible to say how much the remainder have changed. Given the subject matter, some technologies are of far lesser significance than they were two years ago. Given time, most will fade into insignificance.

The system for ordering resources in CoFIND 1 is based on a simple count of votes for selected qualities, accompanied by a bar-graph display of those votes. This becomes unwieldy as more votes are generated, leading to bar graphs which take up most of the available display space, even when using a large, high-resolution monitor. In essence, the system was not scalable, although it did not reach a point of awkwardness during this exercise.

3. Conclusions for the CoFIND 1 study

1. CoFIND 2

Most of the lessons learnt in CoFIND 1 were fed into the design of CoFIND 2, with varying degrees of success. The experiments performed using iterations of CoFIND 2 are here somewhat artificially divided into two sections, dealing separately with topics and qualities, these being the two areas of greatest interest while I developed the system. In real life, the two were usually inseparable and interlinked, which is reflected in references within the text between the two.

2. Topics
1. Islands and bridges

Although it provided a useful method for appraisal of resources, CoFIND 1 was doomed only to be capable of supporting a single subject area and topic, strongly focussed. This was a little like providing an island (to improve the rate of evolution) but failing to provide any means for the island population to meet any others. Two alternative solutions presented themselves. The first (and the more ambitious) was to allow different iterations of CoFIND to talk to each other, exchanging qualities and resources. This raised a range of wonderful possibilities such as the invasion of one ecosystem with another or for cross breeding to occur. However, maintaining separate databases seemed inefficient and hard to control, and suffered from the disadvantage that resources that might be appropriate in more than one topic (for instance, a portal that might be good for everything from weather to shopping) would not be available immediately to users of both systems. A simpler solution would be to keep virtual ecologies within a single system. CoFIND 2 was therefore designed to incorporate the concept of topics, following conventional terminology and the use of the word found in the MDC (Model for Distributed Curriculum) framework proposed by Murray :

Topics can refer to any type of knowledge or learned behaviour, including concepts, facts and procedure, and there are no restrictions on the size or level of granularity of the knowledge they refer to.

1. Topics in CoFIND 2 and their weaknesses

Various methods of organising topics were attempted. A simple list, unless it were ordered by an expert, made it difficult to distinguish an appropriate order to follow the topics, nor did it given any means of identifying groupings and subgroupings of topics. A version which allowed users to optionally specify a "parent" topic when adding a new one was built, but this was not at all in keeping with the general principles of self-organisation which I was aiming to achieve and I dropped it shortly thereafter.

Later iterations of CoFIND 2 implemented a system that put most-used topics at the top of the selectable list, which meant that the most popular topic was likely to be selected. This was limiting in that it failed to capture any relationships between topics (as would be required for the implementation of collaborative trails, for instance) and more than one topic might be appropriate at any given time.

2. Collaborative trails

A solution to the organisation of topics was implemented in the next version of CoFIND (2.5), which was to create a physical workspace providing a blank palette that can be filled with topics. Learners can fill these spaces with topics of their choosing, selecting which of four sectors the topics will appear inside. The choice of four sectors is certainly a piece of sky-hooking but is not altogether arbitrary. If each of the four sectors is filled with seven items of short text of the largest possible font size using a standard browser at a resolution of 1024 x 768 pixels, then the items all just about fit on the screen. The significance of the number seven is not that great, but is based on the work of Miller, who demonstrates that this is approximately the number of items that can easily be assimilated at a glance . There are no constraints on the number of topics which may be entered in any sector, leaving the choice of whether to add more to the users of the system. In the longest study so far, over eight months of continual use by thirty-six students, the maximum number entered in any sector is eight, which perhaps suggests some degree of self-limitation in accordance with Miller’s observations. As each sector fills up, competition sets in. More frequently used topics take up more room by showing their names in a larger font.

1. Seeking a successful algorithm

If the upward and downward pressures are equivalent, then topics of roughly equal popularity jostle to be biggest and oscillate around a similar size and, in their jostling, can actually avoid reaching a successful plateau. The first test iteration of this system, based on a simple algorithm where selection resulted in an increase of two magnitudes of font size for the selected topic and a decrease by one magnitude for the others in that sector, resulted in low local optima being reached when there were many topics in a sector.

The next iteration of this algorithm was responsive to the size of the environment, fitting the evolutionary principle of small, parcellated populations evolving more rapidly. Each time a topic is selected the font size rises by a factor related to the number of topics within that sector (current-rating + number-of-topics-in-the-sector, capped at a maximum of 6) and the font sizes of the other topics in that sector all get smaller by a factor based on the inverse of the number of topics (1/number-of-topics-in-the-sector, capped at a minimum of -3). Unpopular topics eventually get to be very small, though in this version there is no actual death, whilst popular topics tend to remain that way for some time.

The mechanism results in a lot of oscillation when there are few topics in the sector (especially in web browsers which are only capable of displaying three different font sizes using the font tag) but is much more stable when there are many topics. By relating growth and shrinkage to the number of competing topics, successful topics will continue to stand out when there are large populations as effectively as when there are small populations. The net result of this process causes a form of stigmergy, where students are led down popular paths, thus increasing the success of those paths.

3. The system in action

Figure 7.3 shows the results of a group of 82 final year undergraduate students using the system to post resources related to network management over a period of four weeks. The students had been presented with the system as a means of sharing their considerable knowledge with each other and thus to provide a good resource for exam revision. Having already identified the importance of motivation in using CoFIND I made use of the existing structures for motivating students, despite my dislike of the general process.

All of the topics shown except for ‘network magazine’ and ‘network management’ were added by me, although all students were able to add more if they wished. To maintain legibility the maximum deviation possible for a font size is between –3 and +6. The small range of possible values contributes greatly to the dynamics of the system, leading to quite abrupt and calamitous changes with very few interactions. This means that generally useful or topical topics get emphasised very quickly. This in turn means that they are more likely to be used, so the system is both shaped by the interactions of its users and in turn shapes those interactions.



Figure .3 – CoFIND 2.6 topics (network management group)

4. Simon’s Clients and Severs [sic]

Figure 7.4 shows an altogether more interesting example, where a group of 36 MSc Information Systems students (inexperienced users with existing degrees in mainly arts and humanities) were receiving their first introduction to CoFIND in a tutorial session. They had been asked to find resources relating to LANs and add them to CoFIND. However, from the size of its label, Web design appears to be a more interesting topic to the students. Even within a controlled environment, CoFIND was exhibiting self-organising characteristics. It seems that the combination of what was clearly an interesting topic together with the large size of the label Web design attracted more visitors than the allotted topic.



Figure .4 – CoFIND 2.6 topic selection screen showing self-organisation (perhaps)

It is possible that it was the interest factor of topic itself that was solely responsible for the number of clicks on the link rather than the effects of stigmergy. Bearing this in mind it is instructive to look at Figure 7.5, which represents the same database some six months later. It may not be clear from the illustration, but it is worth noting that the system acquired a new livery of colours after complaints from the students, suggesting that the interface is another area ripe for the application of some form of self-organisation.



Figure .5 – CoFIND 2.6 topic selection screen, showing Simon’s clients and severs [sic]

The prominent topics are mostly closely related to the development of the course at the time of this snapshot, with a project on ASP and a strong interest in XML in personal projects, for example, relegating the previously ascendant Web design to a minor role, whilst Reading list has almost disappeared, because (presumably) the students already knew all that they needed to know in that topic. Simon’s Clients and Severs [sic], however, is an anomaly. It was claimed to be a mistake by the student who added the topic, caused by a misunderstanding of how to add some of his own work to the system. In fact, there is good evidence that this was not so, because the student repeatedly clicked the topic label over a period of a few minutes immediately after adding it, thus increasing its relative size and prominence. It is (even when spelt correctly) a topic of only marginal interest in the context of the course. What is fascinating is that the topic, though it later waned, got to be quite popular, visited by twelve different students over two months. This compares with only ten different visitors over a similar period to ASP, a topic which was being covered in class and therefore of greater intrinsic interest. It is tempting to conclude that usage was being driven by stigmergy, kick-started by the student who added the topic in the first place. It was specifically mentioned by a student evaluating the system as being "off-putting that a topic entered in error (Simon’s Clients and Servers) was so prominent."

5. Issues still to be resolved with the topic mechanism

The topic mechanism is faced with the problem that, once a population of topics reaches a certain size, novelty will not get a look in if the ecosystem is already well established. A solution may be the introduction of death to topics. Anomalies such as Simon’s Clients and Severs should be capable of dying, but it might be a bad idea to kill off most topics. Assuming the maximum number of topics supported in a given sector were seven (but the principle applies to almost any arbitrary figure we choose) it is entirely conceivable that more than seven topics might be appropriate to a group’s needs. This is a different problem to that of dying qualities. The death of less useful qualities is far less devastating than the death of topics, as whichever quality is chosen the user will still see resources displayed first which most closely match the topic chosen. If a user is seeking resources related to Ethernet there is no point in displaying those relating to mail servers. An alternative might be to relegate rarely used topics to a second page, or to provide users with the ability to hit a kill button that would eventually result in a topic’s death.

Two main dynamics are available to affect change, the selection of topics and the creation of topics. The rate at which these occur can feed back into the dynamics of the system. If the rate of addition of topics is high, so the ‘death’ rate should be higher, whilst if the rate of use of topics in a sector is high it should limit the mortality. Different sectors might exhibit different characteristic behaviours, which would in turn affect how they are used, which would affect behaviour etc. In keeping with the self-organising principles of the system, these behaviours should be affected by the way that users use the system, leading once again to the control mechanism whose mode of control is affected by what it controls.

A further development of this system should be to allow each sector to compete with other sectors, affecting the space that the other sectors take up on the screen and the consequent size that topic labels may reach. This is in keeping with Kelly’s notions of chunking, Gould’s assertions about island populations and Kauffman’s experiments with NK networks .

Although the topic mechanism allows a path through topics to be dynamically generated, it does not make use of any semantic information to establish relationships between different topics. Topics are traversed ‘blindly’ (although the decisions of individuals to traverse those paths are anything but blind) and so links between them, the structuring espoused by Whelan , does not emerge. It would be useful to capture the history of the paths between topics in some way. These could be represented graphically by trails of greater or lesser prominence based upon the paths actually taken by learners, but this remains an unimplemented feature because it is a far from a perfect solution. As in the case of Collective Mental Maps , actual movements through a body of material are temporally contextualised and do not necessarily show a particular user’s preference given reflective experience, only what they did at a given time (which may well have been mistaken or accidental).

6. Conclusions

• changing font sizes may bring about Stigmergy and evolutionary competition is an effective means of restraint within the system
• the limitation of four sectors is artificial and a clear example of acting as a skyhook. A better mechanism would allow sectors to grow unrestrainedly
• the system as it is does not scale well. With a relatively small number of topics this is partly a problem of interface design, but a large system with unrestrained growth in topics would require the introduction of death or, at the very least, relegation
• there should perhaps be a means of capturing historical information about trails so that future users could benefit from the collaborative trail blazing of earlier cohorts.

1. Qualities
1. Four studies

This group of studies looks at ways in which qualities are used, what benefits and disadvantages they bring and how they affect users of systems. There are four main studies in this section:

Website Evaluation System	This study attempts to look at qualities in isolation in an exercise where students entered and used qualities to evaluate Web sites.
N- vs. 1-dimensional evaluations	This study compares the behaviour of groups who are able to use qualities and groups who are only able to use a single quality.
The effects of topics on qualities	This study compares the use of qualities by the same group but in two different topics.
Stigmergy and qualities	This study looks at the effects of applying two overlapping stigmergic mechanisms to the display of qualities.

It is the nature of these studies that there will often be occasions where the learning outcomes are not necessarily perfectly correlated with the intended object of study.

2. Algorithms for qualities

The POP algorithm used for CoFIND 1 had weaknesses. Popular resources that had been rated early on tended to stay on top, maintaining their advantage even in the face of strong competition. Although this is evidence that there was an effective stigmergic feedback loop (highly rated resources would appear at the top of the list of returned resources, thus making it more likely that they would be selected and thereby increasing the probability of their being rated) the system had a tendency towards a Stalinist regime where new resources were given little chance to develop.

Whilse CoFIND 1’s rating system showed a Stalinist tendency leading to permanently successful resources, by contrast the qualities tended to enter a chaotic regime. Because multiple qualities could easily be selected, there was a high level of dynamism in the movement of qualities on the selectable list.

1. The CoFIND 2 Algorithm

To compensate for the Stalinist tendencies of resource order and to provide scalability, CoFIND 2 uses a six-point Likert scale to rate resources according to qualities, which allows a more sophisticated algorithm to be used to display resources, based on a combination of time-based weighting, averaged ratings and number of votes cast.

To help to avoid the chaotic tendencies of quality order, CoFIND 2‘s selectable qualities were limited to one at a time, selected from a dropdown listbox. In retrospect this bit of interface design was a poor decision, as it led to a single quality (the first on the list) almost invariably being selected, resulting in a Stalinist regime. This was corrected in a lightly later iteration, which displayed the top four qualities in a listbox (not drop-down) without prejudicial pre-selection (Figure 7.6).




Figure .6 – CoFIND 2 resource display screen

Even in this form, the feedback loop is highly stigmergic, and plays a strong role in preventing lower-placed qualities from rising up the list. Little-used qualities provide small lists of relevant resources, so remaining little used. To prevent stigmergy drifting into Stalinism, when new qualities are added they are given a weighting equivalent to the maximum number of votes cast, a system that ensures that all qualities have a chance to survive. This mechanism is a skyhook, but it is a pragmatic solution to the problems of novelty. In a large system (exactly as in large populations in biological ecosystems) there is little chance for novel adaptations to spread.

2. The weighting mechanism in CoFIND 2

New qualities start with a weighting equivalent to a count of the largest number of ratings given to any quality on the system. This usually leads to the new quality being near the top of the list, but seldom right at the top, as existing high-flying qualities will usually have some weighting of their own.

Each time a quality is selected by any user, the weighting for that quality increases by one. Each time a user logs into the system (any user), the weighting is reduced by one for all qualities at once. This process is applied to all qualities equally until they reach a weighting of zero, at which point their success is determined solely by usage. If the system is little used, with few logins, weightings in general will be fairly high whereas a heavily used system will reduce weightings rapidly. This is in contrast to the simple time-based system of CoFIND 1, which led to no qualities surviving if the system was little used. The effect of CoFIND 2’s weighting algorithm is to adjust the dynamics of quality starvation according to the level of usage of the system. Weightings are multiplied with the count of votes to provide rank order. The system is thus organised dynamically according to level of usage.

3. Resource display in CoFIND 2

Resources are displayed in order of the average of their ratings. On its own, this would mean that a resource that has been rated once could be considered as more successful than one that has been rated many times, even though that single rating might be strongly atypical. It would seem that a large number of ratings are more likely to give a reliable indication of the value of a resource than a few. Therefore, to provide a positive skew where more ratings have been given, more frequently rated resources appear before less-rated resources where the average rating is the same. The algorithm to achieve this takes the average rating and added a decimal fraction based on the cosine of the number of votes:

SCORE(A)= ((rating(U(1)+...+rating(U(N))/SUM(U(N)))+COS(SUM(U(N))

There are weaknesses in this system. In particular, there is an averaging of the averages: an average of 4.99 with few ratings could be treated the same as an average of 4.01 with many ratings. It also still means that if ninety nine out of one hundred users rated a resource highly and one rated it low, that resource would appear lower in the list than one that had been rated highly once by a single user. This is not usually as great a problem as it may appear, as a fast-acting negative feedback loop prevents a low-rated resource from staying at the top for long. If a resource were selected because of its early appearance in the list and it turned out to be disappointing, it would probably receive a low vote, thus rapidly reducing the average rating and sending it tumbling down. Similarly, votes for often-rated resources would be less statistically significant than those which were seldom rated. This is in keeping with the way that parcellated populations are able to change rapidly in biological ecosystems and indeed provides concrete evidence that the phenomenon occurs. The larger the population, the less the effects of changes to particular species. This maintains a balance just the ordered side of chaos, and so helps to maintain distinct species, giving sufficient time for a proper ecology to develop.

2. WebSite Evaluation System
1. The problem

Even in the early CoFIND 1 experiments it was hard to separate out the effects of qualities, given the effects of constantly adding new resources. Once topics were introduced, it became even harder to examine qualities in isolation. With the help of my supervisor (Chris Boyne) an experiment was designed to look at qualities in isolation, without giving users the ability to add resources or topics.

2. The process

This study was first discussed in Dron et al . It involved a group of fifty level two undergraduate students at the start of their Human-Computer Interface (HCI) module. The students were all very familiar with using the Web but had not yet studied HCI in any detail.

A two week exercise was prepared which was intended to encourage students to reflect upon the qualities of Web sites which matter when choosing to rate them as good or bad. A version of CoFIND 2 was prepared which did not allow the students to do anything apart from add qualities and use them to rate resources. The objective was to identify how qualities would be used when divorced from dynamically changing lists of resources and topics.

Twenty Web sites were carefully chosen by the tutors of the course to cover a wide range of subjects and HCI issues. All the sites had good points, but the tutors considered that each suffered one or more weakness in their design for usability. These sites were entered into CoFIND by the lead tutor. The students were strongly encouraged to rate the resources using the existing qualities or (should they prove insufficient) to add new qualities. Students were also free to use the discussion mechanism to comment on the resources if they wished, but there was no requirement for them to do so (and none of them did so).

The system was seeded with four qualities (frustrating, hard-to-use, attractive, interesting content), mainly by way of example. These qualities were used by the lead tutor to rate the resources as he saw fit, so that from the start selecting any quality would produce a list of relevant resources. Negative qualities were thus encouraged by this example.

All students were asked to contribute to the system by evaluating at least ten sites and to add new ones if they wished.

3. Evaluation

Three sites received very high ratings: the University of Brighton’s own site, SNARG (http://snarg.net a remarkable Web art installation) and to a lesser extent the site for the Beano comic. The ratings for each of these sites were very mixed and demonstrate that a typical good-bad range of votes of the sort used by a conventional collaborative filter fails to capture the subtlety of feelings that mark out a user’s attitude to a site. For example, SNARG achieved high ratings for positive quality of innovative as well as the negative quality of confusing whilst getting low ratings for hard-to-use and frustrating (i.e. users felt that it was not hard-to-use and not frustrating). Despite being a dedicatedly visual site that presents itself as a work of art, SNARG received low-average ratings for attractive and artistic. It would have been hard to glean this information using traditional two-dimensional numeric ratings and much more difficult to glean a consensus opinion using conventional seals of approval such as abstracts and critiques. Had I provided the students with pre-entered qualities I would not have come up with Author info or Association member, although when presenting these results to a group of Web design teachers they revealed that these were exactly the kinds of things they encouraged students to look for as a guide to the quality of a site. This indication that, even at a relatively low level in the educational hierarchy, the students themselves were able to share sophisticated concepts without tutor intervention can only be regarded as a very positive indicator of the potential for a CoFIND-like system.

Only seventeen of the twenty resources were rated by the students. The three unrated resources were not hidden away at the bottom of the list of the returned resources for any of the originally entered qualities, so it is tempting to conclude that students were put off the sites by their titular descriptions. This most basic, influential and accessible form of metadata is clearly of great importance and raises a wealth of thorny issues already alluded to, especially with regard to the problems of shared classifications. If learners do not share a vocabulary then useful resources might well be overlooked, although the use of ratings at least ensures multiple perspectives on the value of a resource so that resources in CoFIND perhaps stand a better chance than those in systems which lack a rating mechanism.

Only four of the fifteen qualities added by students were negative, and these were the first four added at the start of the assignment. As the assignment progressed, all of the remaining qualities were positive. Although the negative qualities continued to attract a reasonable number of votes (relatively speaking, proportionally the same as those given to positive qualities), there is a suggestion that positive qualities were considered more appropriate when viewing sites. It seems likely that the early arrivals were influenced by the seeding qualities entered by the tutor, suggesting that some reproduction was involved, if only at a high level of abstraction.

As can be seen from Table 7.2, there appears to be a certain amount of stigmergy going on- qualities that were created at the start remained popular throughout, an unlikely event were there not a positive feedback loop driving it. Many qualities made a brief appearance then died, despite the fact that they were given a weighting algorithm to start them off near the top of the list, so stigmergy did not lead to a Stalinist regime.

Table .2 – use of qualities over time for the HCI module

Quality and date added	Total votes	15/02	16/02	17/02	18/02	19/02	21/02	22/02
Frustrating (added 10/02/00 16:06:00)	9	2	1	2	2		2
Hard to use (added 10/02/00 16:13:18)	11	4	2	3	1			1
Attractive (added 10/02/00 16:14:29)	40	9	9	13	7			2
Interesting content (added 10/02/00 16:17:34)	13	6	2	2	2		1
Messy (added 16/02/00 14:33:27)	3		2		1
Slow (added 16/02/00 14:36:31)	5		2	1	1		1
Boring (added 16/02/00 14:41:44)	5		1	3			1
Confusing (added 17/02/00 09:06:18)	10			6	4
Navigation (added 17/02/00 13:40:09)	3			2	1
Orientation (added 17/02/00 13:40:30)	6			5	1
Appropriate Links (added 17/02/00 13:40:56)	6			4	1	1
Innovative (added 17/02/00 14:00:11)	4			3	1
Association Member (added 17/02/00 14:00:56)	5			4	1
Author Info (added 17/02/00 14:09:19)	5			4	1
fun (added 17/02/00 15:10:02)	3			1	1			1
Artistic (added 17/02/00 15:26:52)	4			1	1	1	1
Informative (added 17/02/00 15:35:16)	4			2	1		1
Indispensable (added 19/02/00 15:51:18)	2					1	1
rich (added 21/02/00 13:14:11)	3						3
TOTAL VOTES	141	21	19	56	27	3	11	4

5. Conclusions

• qualities provide useful differentiation in determining the ways that a site can be valuable. Negative qualities are less likely to be entered than positive qualities
• qualities are not the only metadata. The fact that some resources were unrated strongly indicates the importance of names and descriptions in adding resources. This is a problem that is unlikely to go away in any iteration of CoFIND, unless resources are represented by relatively value-free icons or numbers. Given CoFIND’s usual need for a rich learning environment with as many cues as possible to help learners find good resources, this might not be very practical
• stigmergy again proves an important factor in guiding users to relevant resources and the algorithm used does not appear to lead to a Stalinist regime
• the qualities that were entered are related to the use to which they were put. Due to the requirement of the assignment to rate web sites, few of the qualities which were entered related to specific learning needs, but instead captured something about the quality of web sites from a variety of perspectives. This is a point to which will be returned to later.

1. N- vs. 1- dimensional evaluations
1. The problem

Having confirmed that there was some value in the use of qualities, it was worth investigating ways in which the behaviour of users was affected by the ability to use qualities, to use an n-dimensional as opposed to a single-dimensional collaborative filter. I was interested in the effects of qualities on the way that the system was used.

2. Design

A version of CoFIND 2.5 was built with the option to provide an interface for any selected user which did not make explicit use of qualities, but instead provided a simple Likert scale of good to bad. In all other respects (including the manner of rating) the system remained the same. Resources were displayed in order of the average of all ratings, the tentative (but as it happens correct) assumption being that most if not all qualities entered would be positive.

With two broadly comparable groups at my disposal (final year network management students and MSc students studying network technologies) for whom CoFIND would make an appropriate learning aid, I used the same system for both groups, hoping that this would provide some cross-referencing, leading to more valid results.

3. The process

Approximately half of eighty-two students on a final year network management module were randomly assigned by the system to use the simplified, quality-less interface, whilst the remainder could use qualities. Neither group was explicitly informed of the other, although the fact that the system was used in a crowded computer laboratory meant that there was a likelihood that neighbours might spot a different interface.

Users able to use qualities rated resources slightly higher overall than those who were not, although this was not statistically significant. In its first day of use, the system started with four qualities that had been added by me and one that was added by a student ("Topical"). Only one further quality was added over the entire four weeks ("to some extent complex"). Overall, the figures shown in Table 7.3 were achieved after four weeks using the system, by which time there were ninety-one resources available (fifteen added by me, the remainder added by the students).

Table .3 – the effect of the use of qualities on ratings for the network management group- overall

Use of qualities	Average rating	Number of ratings
Using qualities	2.9	66
Not using qualities	2.6	73

 

Breaking the ratings down by quality reveals the pattern of behaviour indicated in Table 7.4.

Table .4 – the effect of use of qualities on ratings for the network management group- by quality

Average rating	Number of ratings	Quality
2.9	41	Good
2.67	3	Informative
3	9	Reliable
2.67	6	Simple
3	2	To some extent Complex
3.4	5	Topical<