Saturday, 5 March 2016

The Cybernetics of Learning in Stafford Beer and Gordon Pask

One of the definitions of Cybernetics is that it is a "Way of thinking", or even a "Way of thinking about ways of thinking" (this is by Larry Richards). As a 'way' the important thing with cybernetic inquiry it seems to me is that it is a method. It's rather like Husserl's phenomenology which, rather than aiming to specify the structure of consciousness (which was later attempted by the likes of Heidegger), Husserl was anxious to specify the method for approaching consciousness. Increasingly I think cybernetics is a method for approaching science. It is meta-scientific.

So then I began thinking about how the classic cybernetic developments of the past can be seen as methods. Of most interest are the two which I have had most to do with in education: Stafford Beer's Viable System Model and Gordon Pask's conversation theory. So here goes:

Stafford Beer’s Management Cybernetics

Stafford Beer’s Management Cybernetics is an approach to learning and communication for people within organisations. At the heart of Beer’s understanding of the cybernetics of management is a concept of ‘control’:
“Control is an attribute of a system. This word is not used in the way in which either an office manager or a gambler might use it; it is used as a name for connectiveness” (Cybernetics and Management).
Control, in other words, is an index of relationships. More specifically, to examine a “control system” – a set of components between which there are relationships – is to examine patterns of constraint where each component is subject to constraints applied from its environment (which is composed of other components).

Beer’s question concerns the specification of types of relationship expressed in models and their presence in nature. He developed a generic model which specified a set of relationships between regulating mechanisms (that is, mechanisms which exercise constraint or control on each other). Drawing on Ashby’s multi-level regulating mechanisms, Beer considers that relationships at macro and micro levels exhibit particular kinds of pattern: that is, there were particular distinctions which could be assigned to different kinds of relationship and that these distinctions could be applied at different levels of a system. Drawing on the human body, he considered the relationship between the different components of brain, heart, lungs and so on, the endocrine system, the circulatory system, and so on. There were relationships between the components which demanded different patterns of coordination. There were different kinds of constraint which needed to be exercised depending on what it was a component was doing and where it was doing it. For example, there were relationships of self-organisation as the adaptation of one entity would coordinate with the adaptation of another entity. There were also relationships between a higher observer and these self-organising entities. Then there were coordinating forces which controlled the behaviour of the components within an environment. Finally, there was an executive function whose role was to balance and coordinate the respective needs of the other components.

Beer considered two fundamental kinds of distinction – the distinction between the different kinds of relationship within a system, and the distinction between levels of recursion between a system's components. There were different regulating systems and different channels of communication. Additionally, there were broader distinctions concerning the degree of coordination which occurred from top to bottom, and the degree of self-organisation which occurred horizontally.

The distinctions of the Viable System Model provide a way for people to talk about the organisational situation they find themselves in. The empirical process of Beer’s approach is one of relating the distinctions of the model to distinctions about reality. This is a stimulus for conversation. Beer’s model encourages people to express their understanding or lack of it in terms of the concepts inherent in the model itself. Since there are few concepts in Beer’s model, the result is an attenuation of the concepts expressed by participants, where key questions like “what are the components of the business?”, “at what level of recursion do those concepts exist?” are asked.

Beer’s model effectively operates as a way of ‘indexing’ real experience: distinctions are made in the model, and those distinctions are identified in key experiences and things in the real world. In the use of the model, what is agreed are the distinctions that connect the shared experience of the environment with the model. In the process of identifying this mapping, something important happens: the kind of mapping between the model and experience of reality is necessarily reductive. There will always be some phenomenon which exists outside the index. It may only be a feeling – but the inability to express it within the context of the present understanding of the model drives a process of deeper inquiry. Such moments of discovery of that which can't be explained within the given model is a moment of identifying constraint. The challenge it presents is to rethink the mapping of experiences on reality to the distinctions of the model.

In thinking through the different ways of applying indexes to reality, different kinds of constraint are revealed. This identification of constraint in the model was fundamental to the processes which drove the conversation. The misidentification of a system component or a level of recursion would gradually emerge in the configuring of the model to the real environment. Beer’s model is richly generative of possibilities. Constraints are identified by exploring the different ways that the model might map on to nature. Yet the locus of constraints is not abstract; it is concrete within the relationships between those who use the model to talk about their experience of the business. Like many other systems management techniques, the VSM is a tool for coordinating understanding, which necessarily involves sharing understanding of the different constraints that different actors in the business operate within. 

Gordon Pask and Education

Gordon Pask was a cybernetician whose areas of innovation were in educational technology, art and music and architecture. Pask’s approach to the cybernetics of learning differed from Beer’s Viable System Model, in the sense that he sought not to identify a specific generative model which could be mapped onto reality, but rather to think about the inter-human conditions within which each of us build our own models as part of our learning. For this reason, Pask significance to educational thinking is based on his assumption that each individual, teachers and learners, is a kind of “cybernetician”, continually building models of the world and asking questions of reality in the light of their models and adapting as their environment changes.

Whilst for Beer, these adaptive processes may be accounted for in the balancing of variety between the different components of VSM, for Pask, the adaptive process arose through the emergence of distinctions in conversation with others and in the shared engagement with the environment. In order to express this, Pask specified a broad description of the organisational situation individuals find themselves in when they engage in discussing their environment and their knowledge. The resulting interaction dynamics share many properties with the dynamics of conversations around the VSM – including Ashby’s concept of multi-level mechanisms - except for the fact that there are no fixed distinctions.

Pask presents what on first inspection looks like a rather cold “computational” view of the human being. He discusses how the adaptive cognitive apparatus of human psychology (what he calls the P-individual) works as ‘software’ running on the ‘hardware’ of the human brain – something he called the M-individual. The conjunctions between a P-individual and an M-individual engage interactively in a process Pask called ‘conversation’ which he imagined was rather like the ‘dance’ between the regulators in Ashby’s homeostat. His view was that:
The real generative processes of the emergence of mind and the production of knowledge can be usefully modelled as multilevel conversations between conversants (some called P-individuals, others merely “participants”) interacting through a modelling and simulation facility.”

The important thing here was the idea of a ‘modelling and simulation facility’. This was, in effect, the negotiating table around which distinctions could be agreed between the teacher and the learner. Of course, the specification of a modelling and simulation facility was also a spur to the creation of a series of highly innovative teaching machines who purpose was to facilitate and explore the ways in which agreements about distinctions could be managed.

Like Ashby and Beer, Pask saw the discursive process as a multi-layered regulatory system. Distinctions existed at different levels with distinct relations to one another: the learning process involved understanding the relations between distinctions in the same way that mapping of the VSM involved the specification of levels of recursion, or the understanding of the difference between ‘system 1’ and ‘system 4’.

Coexisting in the conversation process, there were many different things happening. At one level, there were utterances about distinction, at another, practical engagement with the environment, or coordinating instructions between the learner and the teacher. Pask argues:

“Various emergent levels and meta-levels of command control and query (cybernetic) language (L0 L1—Ln L) need to be explicitly recognized, distinguished, and used in strategically and tactically optimal ways.”

Within Pask’s P-individual there was a coordination of utterances within a conversation - which in various ways interacted with other kinds of discourse. Pask’s saw the P-individual using a computational metaphor: the P-individual was a kind of “algorithmic procedure” enacted on the environment and producing an output which was then processed in other ways. With this computational methaphor, he believed emergent dynamics could give rise to new ideas, and alongside this, a new human actor, team or organisation might emerge. Supporting the computational process were physiological functions located in the ‘hardware’ (brains) performing more fundamental operations like memory. Not surprisingly, this computational account excludes the role emotion plays in conversation: his model considered what he called a “strict conversation model” which ignored the peoples’ feelings, but concentrated on utterances, psychomotor skills and perception.
The theory specifies a set of relationships: there is a relationship between physical systems – the environment and the biological constitution of bodies; and there is a relationship between concepts existing in minds which interact through conversation. Each actor harbours a model – the articulation of set of ideal possibilities. The conversation is driven by the difference between the model in the teacher and the model in the learner. Both the teacher and the learner are building models and then exploring the models for their fit in the environment. As they do so, they encounter phenomena which don’t fit their existing model: for example, something may occur in the environment which they don’t understand, or the teacher (who has a deeper knowledge of concepts) will say something which exists outside the learner’s current model.

What occurs in this process is that errors in the respective models are identified: the constraints or disparity between ideas and experience. Something may occur which demands that the learner has to find a new articulation of some constraint. The distinctions which the learner has to coordinate are concepts, and the reaction to events is to reorganise, or re-express concepts: both indexing concepts with reality, or seeing concepts in a different kind of hierarchical relation to one another. This necessitated that the articulated concepts had to have some kind of structure: some concepts were more universal or deeper than others. Discovering the hierarchy of concepts was a way of reframing indexes of concepts and positions of concepts. Where Beer’s concepts caused a simplified conversation of calibrating a rich and powerful generative model with reality, Pask’s conceptual positioned complex systems (people) in a way where the logical possibilities of the model were generated through the process of conversation. This had the advantage of not requiring a specialised language: a concept is a concept in whichever environment it occurs. However, Pask’s model assumed much about the constraints which would be experienced by learners, and the reasons for those constraints would be lack of understanding about concepts. What Pask didn’t consider was that the constraints bearing upon learners were emotional. This is particularly apparent in the domain of educational technology.

Pask’s conversation theory is both a theory about learning and adaptation and a meta-description of what it is to do cybernetics. We might, for example, imagine that Ashby has a Paskian conversation with his model of the brain, and that in the light of his experience with reality, he pursued errors in his distinctions and generated of new concepts. Equally, Beer’s setting upon the Viable System Model was a way of setting upon reality and making distinctions about it where applying it led to the identification of deficiencies in particular mappings, and the search for better mappings.

In real education, there is no learning interaction where emotion does not play a major role. This would suggest that Pask’s concentration on computation at the expense of emotion in conversation is deeply deficient. However, emotional blockages are precisely the kind of ‘error’ which might be identified in a Paskian learning conversation. More fundamentally still, however, is the problem as to whether the pursuit of error is something that learners and teachers actually do in their conversations. Whilst Pask’s computational model presents the pursuit of error as rational, it is often the constraints of emotion which prevent individuals behaving in such rational error-seeking ways. However, the value of Pask’s model is that it lends itself to educational intervention, and it is through educational intervention where the strengths and weaknesses of the model can be explored.

Pask himself developed a number of teaching machines in the 1950s. Mostly these concerned the teaching of well-defined cognitive skills such as the programming of ‘punch cards’ (the main means of data input into the computers of the time). In 1999, Pask’s model became the foundation of a broader ‘conversation model’ in educational technology, promoted by Diana Laurillard. Laurillard’s presentation of the conversation left out much of the detail of the broader theory behind Pask’s work. Yet, to its audience – who were then enthralled and threatened by fast approaching technology - this did not matter: it was clear that computers supported new ways of conducting conversations, and that within the dynamics of those conversations, there were processes of ‘teach back’ between teachers and learners working in a shared environment. Yet the nature of the learning environment was complex.

These complex inner dynamics could, of course, be ignored in the general presentation of the interaction between learners and technology. For example, Boyd argues that the conversation model fits the following situation:

"A is a medical student and B is an engineering student. The modeling facility they have to work with might be Pask’s CASTE (Course Assembly System and Tutorial Environment, Pask,1975); equally possibly now one might prefer STELLA or prepared workspaces based on Maple, MathCad, or Jaworski’s j-Maps. The recording and playback system may conveniently be on the same computers as the modeling facility, and can keep track of everything done and said, very systematically.”

Boyd illustrates the kinds of conversation that might follow.
“In reply to some question by A such as, “HOW do engineers make closed loop control work without ‘hunting’?” B acts on the modelling facility to choose a model and set it running as a simulation. At the same time B explains to A how B is doing this. They both observe what is going on and what the graph of the systems behaviour over time looks like. A asks B, “WHY does it oscillate like that?” B explains to A, “BECAUSE of the negative feedback loop parameters we put in.” Then from the other perspective B asks A, “How do you model locomotor ataxia?” A sets up a model of that in STELLA and explains How A chose the variables used. After running simulations on that model, A and B discuss WHY it works that way, and HOW it is similar to the engineering example, and HOW and WHY they differ. And so on and on until they both agree about what generates the activity, and why, and what everything should be called." 

Within this situation, Boyd argues that t is possible to determine different levels of (Ashbian) regulation occurring within each learner.
“Level 0–Both participants are doing some actions in, say, CASTE (or, say, STELLATM), and observing results (with, say, THOUGHTSTICKER) all the while noting the actions and the results.
Level 1—The participants are naming and stating WHAT action is being done, and what is observed, to each other (and to THOUGHTSTICKER, possibly positioned as a computer mediated communication interface between them).
Level 2—They are asking and explaining WHY to each other, learning why it works.
Level 3—Methodological discussion about why particular explanatory/predictive models were and are chosen, why particular simulation parameters are changed, etc..
Level 4—When necessary the participants are trying to figure out WHY unexpected results actually occurred, by consulting (THOUGHTSTICKER and) each other to debug their own thinking.”

This is clearly something of an ideal situation. In reality, as with anything to do with technology, there are a whole set of unarticulated conversaations and levels of communication which can create within the participants of a conversation various kinds of emotional confusion. Whilst it may be considered that the conversation between scientists is a multi-level emergent discovery of convpts at different levels, the conversation between a teacher and a learner is much more complex. Most fundamentally, the ‘why’ questions which are asked by the teacher are not the why questions asked by the learner: they are why questions asked about the learner. Fundamentally, what constraints are producing the behaviour/utterances that are witnessed?

In terms of characterising these constraints, the most important element is the systemic multi-level constraints which might produce the kind of emotional confusion which was articulated by Bateson in his ‘double-bind’ theory. The particular problem of technological engagement is that the technology its itself a constraint of the communication of constraint. Moreover, it is a constraint bearing upon the communication of constraint whose position in the conversation is asserted by a powerful figure (say, a learning technologist, or an institution).

Given the identification of emotional constraint, is to think of ways in which the deficiencies in the model might be addressed. Here it may be argued that Pask failed to see how the different levels of conversation get tied up with one another. Bateson on the other hand, was particularly interested in this aspect. Watslawick explains this dimension: Watslawick explains that:
"Once it is realized that statements cannot always be taken at face value, least of all in the presence of psychopathology - that people can very well say something and mean something else - and, [...] that there are questions the answers to which may be totally outside their awareness, then the need for different approaches becomes obvious."
Saying and meaning are very different things. The latter has to do with expectations and is a much deeper level function. Watslawick quotes Bateson saying:
"as we go up the scale of orders of learning, we come into regions of more and more abstract patterning, which are less and less subject to conscious inspection. The more abstract - the more general and formal the premises upon which we put our patterns together - the more deeply sunk these are in the neurological or psychological levels and the less accessible they are to conscious control."

Is is then enough to include the double-binds that might interfere between different levels of regulation and then to think what it is the teacher should do in response. The issue is that fundamentally Pask’s thesis that communication is a coordination of coordinating mechanisms, the utterances of words and their comparison by a teachers is too shallow. Indeed, it fails to match what Pask himself was doing in his exploration of conversation. Pask must have observed and participated in conversations and wondered what they were about. But what if the model Pask sought was the same process that Pask himself went through in coming up with the theory.


There are many definitions of cybernetics but behind them all is what amounts to a way of thinking which is distinct from the approach of classical science. At the heart of this “strangeness” is the idea that cybernetics does not seek to uncover causal relations. Whilst it has its own body of empirical practices involving the building of models and machines, it does so to explore the relations between the generative possibilities of ideas about how the world works, and how nature appears to work. Cybernetics fundamentally orients around the pursuit of error in the relations between ideas and nature, not the search for proof of ideas: it can be thought of as in a continuous process of falsification. Cybernetics makes the appeal for its approach by arguing that the problems of the world, from the ecological catastrophe, to social inequality and political turmoil result from what Bateson calls a “disparity between the way nature works and the way humans think”. It therefore can make the claim that the best way of addressing this disparity is to turn it itself into the object of scientific investigation.
In order to identify and agree error, it is important to have a shared understanding of whatever model or theory is being compared to nature. The value of any theory is not its particular explanatory power, but that it can specify mechanisms which can be agreed by a group of scientists or educationalists. With agreement over a clearly-expressed theory, attention can focus on practice where attempts can be made to map the distinctions of theory to nature. Because theories are deficient, there will be agreement and disagreement regarding the interpretation of things that happen. Where what happens does not map to the model, or what is predicted in the model does not happen, then cybernetic science moves forwards.
I have tried to explain how this fundamental methodological orientation can characterise many cybernetic interventions in the past. The two cases we considered took a slightly different path, but deep down they are the same. Beer’s approach was to produce a model with various distinctions, and for individuals to explore their experience of their organisational environment by agreeing a map of these distinctions to their understanding of the organisation. Pask’s conversation theory comes with no predefined distinctions of this sort, but rather a description of inter-human dynamics which would generate distinctions which would then be agreed as part of a learning process as the conversation evolves in a shared environment.
In the reality of education, we do both. Syllabi and curricula present predefined distinctions (albeit not as rich and powerful as Beer’s) to which a learner’s exploration of their learning material must eventually fit. At the same time, conversations are the driver for the emergence of new distinctions whose disparity with experience can either be the cause for further and deeper inquiry and discovery, or it can be the driver for the kind of double-bind dynamics which place learners in intractable emotional states of confusion and inaction.
However, so far we have only considered the ways in which the dynamics of a model may be specified and agreed. We have not yet explored the ways in which the experience of nature may be measured and recorded, and those measurements may be related to the dynamics of a model. To do this, we need to explore the dynamics of our relations with nature: the information environment with which our senses interact.

No comments: