Track:
Learning Feature Relevance: Empirical and Forman Analyses
Downloads:
Abstract:
The usual case-based reasoning approach assumes that for each given problem instance it is necessary to retrieve from scratch a similar case from the case base. Therefore, an indexed memory stmctare or other means of facilitating fast access is typically needed. Moreover, a complete solution is usually stored together with each case, that can be adapted to the given problem. We developed a different approach for multi-step problems. It utilizes the information about the relevant case for the last step to quickly find the appropriate case for the current step from only few relevant cases that are connected. Therefore, no special indexing schemata are required. Instead, we store a value for each case and similarity links to other cases, but no solution. For situations outside the scope of the case base we integrated case-based reasoning in several ways with heuristic search. We performed experiments in a game domain, that showed the usefulness of our approach. In particular, we achieved a statistically significant improvement through combination of case-based reasoning with search over pure search or pure casebased reasoning. For multi-step problems, our approach appears to be more useful than the standard approach to case-based reasoning.