AI according to Schank

Roger Schank's group: some of the best-known and longest-running research on symbolic models of memory. What does he think AI is? I've taken this list from his article What is AI anyway? in The foundations of artificial intelligence edited by Partridge and Wilks (CUP 1990; PSY KH:P 025). Many of the comments and examples are mine, not his.

• Representation. Probably the most significant issue. What do we know? How can we get a machine to know it?

  Although Schank doesn't say so, there's more than one aspect to this. E.g. in Poetic, it is not enough to say that the world model uses logic as a knowledge representation. This describes the underlying language, but not what is said in it. We need to consider both the symbol level (elementary symbols and operations), and the knowledge level (description in terms of goals etc). See entry for Knowledge level in the Encyclopaedia of AI.

• Indexing. See the paradox of the expert, last week. Note the emphasis on indexing in the memory model I describe below.

  Important point: we should not constrain our engineering or our cognitive models by what conventional computers can do. As connectionism suggests, biological hardware is very different. It may for example be more efficient at ``automatically'' performing associations that, on a conventional machine, would require complex indexing methods. For an extreme example, the holographic pattern matcher described in Transforming the prospects for robot vision, AI photocopy B165.

• Dynamic modification. Any intelligent program will need to change its methods of representation. For example, memory for chess positions, last week.

• Decoding from the real world into an internal representation. E.g. Poetic going from police logs to its logic-based world model. How do we decode, how do we cross-reference sensory data, and so on? Indeed, do we do so at all?

• Inference. How do we combine pieces of information to make new ones? And when? For example, on reading ``John went down the aisle and put a can of tuna in his basket'', do we infer then that he was in a supermarket, or do we wait until we need to know what he was doing?

  The issue of timing is relevant to generalisation (see below). Do we store examples and generalise later, or is a certain amount of generalisation automatically performed as we store each example?

• Controlling the combinatorial explosion. How do you prevent inference going on for ever? How do you decide how much you want to know? See the fable about R2D2 in Cognitive Wheels by Dennett, AI box photocopy D74.

• Generalization. A good generalizer must be able to connect disparate experiences (the essence of creativity, says Schank). An excellent example: the Bongard problems in Chapter XIX of Gödel, Escher, Bach by Hofstadter (PSY KH:H 067). Here, generalisation is finding what the figures on each side have in common, that distinguishes them from those on the other side. But the connection may not fit further examples (problem of induction), so the generaliser must be able to experiment, re-fit and revise.

• Curiosity. In Schank's view, curiosity depends on prediction. A system's predictions may fail, and it should try to find out why. This will involve formulating suitable questions and ways to answer them - the answers might come from internal knowledge, or by experiment. These predictions might arise from generalisation, but could also come from other kinds of processing: e.g. a plan that fails.

• Prediction and recovery. Before you can discover why a prediction failed, you must be able to tell that it has failed.

• Creativity. There are very few creative AI programs, AM being one exception. Briefly, a program that created new mathematical concepts and conjectures from old, starting from set theory. It didn't churn them out at random (not just throwing any ideas together), but ranked them by interestingness, determined by (e.g.) how often the same concept had been discovered; how closely it was related to other interesting concepts.