Learning to improve case adaptation

Case-based reasoning (CBR) solves new problems by retrieving records of similar past problem solving episodes and adapting the prior solutions to t the current situation. While the retrieval phase of CBR has been explored with success by past models, developing e ective algorithms for automated adaptation remains an open problem. The central hypothesis of this research is that e ective case adaptation knowledge can be learned and reapplied automatically by applying CBR to the adaptation process. In this model, adaptation knowledge is learned by storing the results of successful adaptations to be reused in solving similar future problems. If there is no relevant adaptation knowledge to reuse, general rule-based methods of adaptation are used to build the adaptation case base. These methods model the search for needed information as a planful process whose strategies can be captured and reused by case-based reasoning when similar situations are encountered in the future. In addition, as adaptation knowledge is acquired, methods for evaluating the similarity of past cases are re ned to re ect the adaptability of a prior case to the new situation. This model is implemented in the DIAL system, a case-based planner in the domain of disaster response planning. DIAL analyzes new disaster situations and proposes response plans to address the problems arising from the disaster. In this research, two criteria are used to evaluate adaptation learning in the DIAL system: the eÆciency of the solution process and the usefulness of its results. The eÆciency of the solution process is examined through statistical evaluation of empirical results. Usefulness is de ned as the system's ability to generate acceptable solutions. Through analysis of the results, the utility of this approach is measured and the contribution of the model is judged.