A subsumption architecture for theorem proving?

Brooks has criticized traditional approaches to artificial intelligence as too inefficient. In particular, he has singled out techniques involving search as inadequate to achieve the fast reaction times required by robots and other AI products that need to work in the real world. Instead he proposes the subsumption architecture as an overall organizing principle. This consists of layers of behavioural modules, each of which is capable of carrying out a complete (usually simple) task. He has employed this architecture to build a series of simple mobile robots, but he claims that it is appropriate for all AI products. Brooks’s proposal is usually seen as an example of nouvelle AI, in contrast to good old-fashioned AI (GOFAl). Automatic theorem proving is the archetypal example of GOFAl. The resolution theorem proving technique once served as the engine of AI. Of all areas of AI it seems the most difficult to implement using Brooks’s ideas. It would thus serve as a keen test of Brooks’s proposal to explore to what extent the task of theorem proving can be achieved by a subsumption architecture. Tactics are programs for guiding a theorem prover. They were introduced as an efficient alternative to search-based techniques. In this paper I compare recent work on tactic-based theorem proving with Brooks’s proposals and show that, surprisingly, there is a similarity between them. It thus seems that the distinction between nouvelle AI and GOFAl is not so great as is sometimes claimed. However, this exercise also identifies some criticisms of Brooks’s proposal.