Search in AI: Escaping from the CSP Straightjacket

We investigate some useful strategies for solving a variety of hard search problems. In the process we identify weaknesses in the standard CSP formalisation of such problems. 1 Search in Artificial Intelligence A basic challenge in Artificial Intelligence is to understand how people solve problems. In 1968 Herbert Simon wrote [14]: Problem solving is often described as a search through a vast maze of possibilities...Successful problem solving involves searching that maze selectively and reducing it to manageable proportions Simon suggested that people solve problems by following search strategies. Ultimately, the study of human problem solving reduces to the study of search strategies and how they are developed and applied: Once the strategy is selected, the course of search depends only on the structure of the problem, not on any characteristics of the problem solver. Since 1968 there has been a great deal of investigation into search strategies. The reason has been that many tough problems in AI and other disciplines can only be solved by search, and without a good search strategy these problems could not be solved at all (in any useful time span). What have emerged are a variety of techniques that can be combined into search strategies tuned to specific problem classes. One set of techniques, for example, are those associated with finite domains, including forward checking and fail first. Simon used as an example the crypt-arithmetic problem DONALD + GERALD = ROBERT. He described in detail the set of choices which might be pursued by a thinking person in solving the problem. It is truly remarkable to discover how closely his description fits the behaviour of a constraint program solving this problem using the very techniques mentioned above: forward checking and fail first. In the following we will explore search strategies for solving hard problems. The challenge is no longer simply to capture the “intelligent” behaviour people use in problem solving. Now researchers are developing and applying more and more sophisticated strategies, so that we can program computers to produce answers to problems that were previously beyond our reach. 1 ICL and IC-Parc, Imperial College, London SW7 2AZ, UK, email: [email protected] 2 Formalising Search Some Shaky Foundations