Automatic representation changes in problem solving

We explore methods for improving the performance of AI problem-solving systems by changing problem representations. Researchers have accumulated much evidence of the importance of appropriate representations for the efficiency of AI systems. The same problem may be easy or difficult, depending on the way we describe it and on the search algorithm that we use. Previous work on the automatic improvement of representations has mostly been limited to the design of individual learning algorithms. The user has traditionally been responsible for the choice of algorithms appropriate for a given problem. We describe a system that integrates multiple representation-changing and problem-solving algorithms. The purpose of our research is to formalize the concept of representation, explore its role in problem solving, and confirm the following general hypothesis: An effective representation-changing system can be constructed out of three parts: (1) a library of problem-solving algorithms; (2) a library of algorithms that improve problem representations, by static analysis and learning; (3) a top-level control module that selects appropriate representation changers and a problem solver for each given problem. We have supported this hypothesis by building a system that changes representations in the PRODIGY problem-solving architecture. The library of problem solvers consists of several search engines available in the PRODIGY architecture. The library of representation changers includes novel algorithms for selecting primary effects, generating abstractions, and discarding irrelevant elements of the domain encoding. The control module chooses and applies appropriate representation changers, stores and reuses new representations, and selects problem solvers. To identify effective algorithms, it performs statistical analysis of the past results, as well as heuristic comparisons of the available representations.