litceysel.ru
добавить свой файл
  1 2 3 4 5 6

Dynamics: minds as processes

"a single system (mind), produces all aspects of behaviour. … Even if the mind has parts, modules, components, or whatever, they all mesh together to produce behaviour.... If a theory covers only one part or components, it flirts with trouble from the start. It goes without saying that there are dissociations, independencies, impenetrabilities and modularities. ... but they don't remove the necessity of the theory that provides the total picture and explains the role of the parts."

A. Newell Unified Theories of Cognition, 1992 pages 17-18

At about the same time that Broadbent published Perception and Communication the new field of artificial intelligence was getting going. Two influential conferences took place around this time; the famous 1956 Dartmouth conference where the field of artificial intelligence acquired its name (see http://livinginternet.com/i/ii_ai.htm) and the "Mechanisation of thought" conference in Teddington, England (1957). These meetings brought together a range of scientists who were interested in understanding diverse aspects of mind, from vision and pattern recognition to speech, language and problem solving.

Many in the early AI community thought they were simply developing a new branch of engineering but others felt that their work was shedding light on the general principles that underlie intelligence. In the following decade interest in AI gathered momentum in parallel with the growth of “information processing psychology”. In 1967 the psychologist Ulric Neisser published Cognitive Psychology, a book which brought the static images of cognition that Broadbent and others were using together with ideas emerging from work on computational systems for carrying out complex tasks such as playing board games, interpreting pictures and understanding natural language. Neisser believed that computer algorithms might be the key to understanding how static cognitive models could be extended to explain the dynamic processes that generate behaviour.


Allen Newell (1927-1992) was a computer scientist who attended the Dartmouth and Teddington conferences. With his collaborator Herbert Simon he pioneered the use of computational models in cognitive science. Newell was particularly interested in the dynamic aspects of cognitive processes and how they are controlled and his ambition was to develop a "unified theory of cognition". Until his death in 1992 he worked on a series of simple, elegant and increasingly ambitious systems for modelling mental processes. Through PSG (Newell, 1973), the OPS series of systems (Forgy & McDermott, 1977) to the SOAR cognitive modelling system which embodied what Newell believed to be a general theory of intelligence (Laird et al, 1987; Newell 1992) he progressively refined three fundamental propositions.


  1. In order to understand mind we must separate the fixed architecture from the variable component of cognition. The fixed architecture is the information processing machinery to be found in all human heads; the variable component is, roughly, what we call knowledge.

  2. Newell took the view that understanding knowledge is basic to understanding cognition. His research position was that all knowledge can be represented by simple condition-action rules.

  3. The control structure, the mechanism by which the static cognitive architecture applies knowledge in reasoning, problem solving, decision-making and other cognitive tasks, is a fundamental problem for understanding the nature of a functioning mind.

The fixed architecture that Newell adopted owed much to the model that had been developed by Broadbent and other psychologists. His knowledge representation, the condition-action rule, is a simple yet very general construct that can be used for many purposes, from standard logical operations like deduction to executing actions and plans.2 The dynamic control structure that Newell thought was fundamental to flexible cognition is a cyclical process of selecting and firing rules driven by situations and events, the “recognise-act” cycle.


Newell and his many students explored a range of questions raised by trying to understand cognition in this way. Many different schemes for controlling rule execution were explored, culminating in the SOAR system that combined the recognise-act cycle with goal-oriented problem solving and the ability to learn from experience by acquiring new rules. As time has moved on this particular embodiment of his ideas has become less convincing as it failed in its original form to explain important phenomena of human cognition, for example (Cooper & Shallice, 1995). However, Newell’s line of work has continued to exert influence on psychologists who wish to understand the dynamic aspects of human cognition, for example ACT-R (Anderson & Lebiere, 1998) and 4CAPS (Just, Carpenter & Varma, 1999). Their work combines the theory of production rules with strong empirical research programmes in cognitive psychology and cognitive neuroscience. From the point of view of this discussion the direction that Newell initiated continues to provide a promising view of the human mind that integrates static and dynamic components within a psychologically plausible framework.

During the period that Newell was developing the rule based approach to cognitive systems a more formal approach to AI was also emerging, particularly in Europe. Logic programming was emerging as another important example of rule-based computation in which cognition is modelled as a form of logical reasoning. In contrast to the production rules of OPS and SOAR that would fire in response to situations or events, rules in a logic programming system are evaluated from the point of view of achieving goals. Where production rules lead to a “forward chaining” cycle, logic programming involves a "backward chaining" process of recursive proofs. As in the production rule approach logic programming also viewed control as a fundamental aspect of computation. Indeed Robert Kowalski, one of the founders of logic programming, viewed all of computation as the controlled interpretation of logical expressions, which he famously summarised as algorithm = logic + control (Kowalski, 1979). Logic programming continues to be an influential approach to AI, providing a versatile and expressive programming technique whose mathematical foundations are well understood. For this reason logic and logic programming will continue to play important role in the emergence of formal theories of intelligent systems, for example (Wooldridge, 2000; Fox et al, 2003).


Many ideas in cognitive dynamics have been investigated. Rule-based information processing possibly represents the most flexible single mechanism for implementing cognitive processes that has been identified to date but others will no doubt emerge as cognitive systems probably require different control regimes in different situations. This is explicit in current work on "autonomous agents" where it is recognised that agents need to be able to operate both “reactively”, in response to events, and “deliberatively”, in a purposive manner (Fox et al, 2003).

The COGENT cognitive modelling system described above is not just a tool for creating static images of mental processes; it also incorporates a programming system for constructing and testing models. Cooper (2002) discusses models of reasoning, problem solving, decision-making and language processing for example. The COGENT programming system is a hybrid logic programming and production rule system (Hajnal et al, 1989). Each process in a COGENT model can operate in a reactive or deliberative fashion, or in a mixed mode, depending on the function being modelled. COGENT is a general tool, not a specific theory of cognition like SOAR or ACT-R. It provides a versatile set of tools for modelling cognitive architectures, and simulating their behaviour. Furthermore, COGENT incorporates standard component interfaces and may offer a new way in which cognitive scientists can collaborate: independently developing theories of cognitive functions and connecting them together for testing over the internet.3



<< предыдущая страница   следующая страница >>