Introduction - Soar

Last week I talked about the ``classical'' production system models. This week, the latest and most ambitious: Allen Newell's Soar. This is the culmination of Newell's interests (he died in 1992), and one can see in it the themes that were foreshadowed in his earlier work - described, for example, in You can't play 20 questions with nature and win, in Visual Information Processing, edited by Chase (PSY BC:C 39).

What's the place of Soar in AI and psychology? For a good two-page summary, see the Soar article in the Encyclopaedia of AI. Main points:

• Original goal of the research: to investigate an AI architecture that used problem spaces for solving all its tasks. In terms of the definition of AI that I gave in the first lecture, this would be ``(b) the theoretical study of possible intelligences, and the general principles underlying them''. If you believe the article, Soar wasn't intended as a psychological theory at that stage. However, it was created by two of Newell's students, so the clean separation of historical stages I've cited here may not be entirely correct.

• These aims were later broadened to
  • An integrated intelligent system. I.e. a system that can do all the things you'd expect expect of an AI system: knowledge representation, problem-solving, planning, learning, natural language understanding, ability to cope with dynamic environments. In contrast to most AI programs, which explore only one or two of these competances/mechanisms, but not whether they can be combined with the others. Cf page 305 of 20 questions.

  • A unified theory of cognition. In contrast, most psychological theories consider only a narrow range of phenomena. Cf page 284 of 20 questions.

• Following the last point, note that several of Soar's mechanisms originate in psychology: problem spaces, production systems, chunking. (And note also that all these have been earlier interests of Newell.)

• At this stage of our knowledge, it seems there are many possible kinds of functional architecture for a general intelligence. The main assumption behind Soar is that its architecture is the best.

• Has now been used to model many phenomena: language comprehension, visual attention, planning, attention-switching in driving, as well as ``classical'' puzzle problems.

  I have not looked at these models in detail, but some critics claim that they are too ad hoc. Soar is Turing-complete (see next lecture): in other words, there's a way to program it to do any task that any other programming language (or computer) could do. So, say the critics, it's certainly possible to program planning, language, etc. into Soar: but these programs do not embody enough psychological constraints to act as models of our mental processes.