Using Reinforcement Learning for Similarity Assessment in Case-Based Systems

be a problem when applying CBR to weak-theoretic domains.1 The knowledge elicitation bottleneck—the inability to precisely encode the knowledge used by human experts—is a concern in many knowledge-based applications. Although researchers cite this bottleneck as a justification for CBR techniques,2 use of domain knowledge in indexing means that CBR techniques are not immune to it. We’ve developed reinforcement-trained casebased reasoning, a reinforcement-learning (RL) technique that uses feedback from the environment to learn case similarity. RETCBR expands the domains in which researchers can successfully apply CBR techniques because it requires knowledge only for case recognition, not to determine the best indexing strategies. We’ve implemented two RETCBR similarity assessment methods and tested them in a weatherforecasting application.