Problems with classical (symbolic AI) models

• Performance is in limited, often highly restricted domains - consisting of specific problem situations or microworlds.

• The brittleness problem: SHRDLU's performance breaks down when confronted with an utterance it is not explicitly programmed to handle. Winograd says this is true of any progam whose knowledge of the world is represented in terms of explicit, static rules. Human cognition is flexible, faced with an unsolvable problem, it can come up with a partial solution. Secondly, humans learn from past mistakes.

• The AI models for human cognition were ``ludicrously un(der?)determined'' by the data (Dennett, 1984?). Any number of widely-variant models could equally well model the data.

• Role of logic: Human cognition does not depend exclusively on deductive logic. That would make even ordinary tasks impossible. Heuristics (to overcome combinatorial explosion) is still rule-based. Human cognition uses mechanisms (similar to Wittgenstein's family resemblences) like association and content-based retrieval.

• Common-sense knowledge: The countless tiny facts gathered by us in the course of living.

  Hubert Dreyfus What Computers (Still) Cannot Do maintains that intelligence necessarily requires general commonsense. Eg. Schank's program using a restaurant script might be able to answer, What food was ordered? or, How much did the person pay for his food? But it could not answer What part of the body was used to eat the food? or How many pairs of chopsticks were required to eat the food? So much of life requires us to use our commonsense that we never notice it.

  • There's too much of CS knowledge: too ambitious to try to code it into a computer.

  • CS knowledge is tacit, not propositional (eg. riding a bicycle).

  • Often procedural, not declarative.

  • Derived directly from the richness experiences and so not obtainable without them.

  • Even if it were possible to code CS, the performance of a serial system would get slower. Humans on the other hand get faster when able to exploit additional constraints (Rumelhart, PDP).

Later work in symbolic AI attempted to overcome these problems using more powerful or more efficient architectures within the rule-based paradigm. (Self-modifying production systems, Case-based reasoning)

The SOAR program (State, Operator And Result) (Newell et. al.). Problem solving with a different approach - a lot of domain-specific knowledge, including models and production rules, are in a ``long-term memory.''

When a current situation in the ``working memory'' matches a rule in the LTM, the rule is brought into WM. Beginning with some explicit rules the system ``learns'' new rules and stores them by a ``chunking'' mechanism.

ACT (Adaptive Control of Thought) John Anderson, 1983: A production system model supported by declarative and procedural knowledge structures. (Similar to SOAR?) See discusion in Gardner, 1985, pp. 131-132. See also sec.

CYC Project, Douglas Lenat at MCC Austin, Texas: Estimate of 500,000 commonsense rules now revised to 2 million, to be completed by 2025.