The functional architecture of cognition

This view is often expressed by saying that the the neurophysiological hardware of the brain implements a ``cognitive processor'', a simple symbol-manipulating machine, and that this is a production system. For a statement of this view, see Foundations of Cognitive Science edited by Posner (Springer-Verlag 1989; PSY BH:P 084), Chapter 3.

But there can be many different types of production system. Here's a list of some of the ways they could differ:

1. The symbols allowed - what, and how many?
2. The resolution methods. You will see below that Eden has various other methods I haven't yet mentioned.
3. The capacity of STM. Is there a maximum size, and if so, what happens when more symbols are added? Do old ones drop out? If so, how does this depend on the time they were added? Is there any kind of ``chunking'' mechanism, so that symbols can be grouped into a single slot?
4. The rules' conditions. Can these test perceptual data directly, or must they go via STM? What is the maximum size of pattern you can have? Can you have ``wild card'' patterns - so that _ for example, might match any set of symbols starting with and ending in . Can the conditions check how many of some symbol there are?
5. The rule's actions. Can these generate actions directly, or must these go via STM? How many symbols can they add? Can they delete symbols?

Different theorists have proposed different types of cognitive processor. The ways in which one differs from another: the complexity of its rules, the capacity of STM and so on, these define its functional architecture. In the same way, a VAX, a PC, and an Apple Mac are all computers, with the same general design. But the details of their functional architecture - the maximum size of number that can be stored, the memory capacity, and so on - are all different.

For those who want to follow up such topics, there is a good account of how production systems work in pages 190-199 of the Handbook of AI volume I, by Barr and Feigenbaum (Kaufmann 1981; PSY KH:B 27). There are some famous production-system models of children's development: these are introduced in Computers and Thought - a practical introduction to AI by Sharples, Hogg, Hutchinson, Torrance and Young (MIT press, 1989): PSY KH:S531. The start of chapter 7 places production systems in a psychological context, while pages 226-231 in chapter 8 describe their use in modelling childrens' subtraction. This is also described in Learning and Problem Solving 3, Open University Course D303, Block 4, Units 26-28 (in the Oversize section, AA:O 2-O M), and in the Young and O'Shea article in Cognitive Science for 1981. Note: there is a misprint in Computers and Thought that sets this article in 1982.

Why did Newell and Simon consider production systems good theories? There's an early article by Newell, detailing some of the flaws he finds in experimental psychology, and what production systems have to offer. This appears in You can't play 20 questions with nature and win, in Visual Information Processing, edited by Chase (PSY BC:C 39).

Another view is given in Richard Young's chapter on Production Systems for Modelling Human Cognition in Expert systems in the micro-electronic age edited by Michie (E.U.P. 1980; PSY KH:M 58). One of the most recent approaches is Newell's SOAR, in Unified Theories of Human Cognition by (Harvard 1990; PSY BH:N 044). You will find a short survey of the book in Behavioural and Brain Sciences volume 15, number 3, September 1992. This also contains a number of criticisms of SOAR and the idea of unified theories.