Automatic Discovery of Heuristics for Nondeterministic Programs from Sample Execution Traces

During the last few years a number of relatively effective Artificial Intelligence (AI) programs have been written incorporating considerable amounts of problem specific knowledge. In particular, declarative representations have attracted much attention partly because of the relative ease with which knowledge can be communicated in this form. Unfortunately, straightforward implementation of declaratively specified knowledge corresponds to a nondeterministic program which incurs enormous computational costs. This thesis investigates one way to limit this cost. We develop control heuristics for a family of problems from traces of sample solutions generated during a training session with a human expert. Algorithms are presented which recognize a set of patterns in the sequence of knowledge applications and which compile descriptions of these patterns in a control language, called CRAPS. More specifically, patterns of repeating, parallel and common sequences are considered in the analysis. The analysis also produces a set of meta-rules which aid the CRAPS description in the event the sequencing it specifies is inappropriate. The CRAPS description and meta-rules are then used for guidance in solving subsequent problems.