Case Adaptation with Neural Networks: Capabilities and Limitations

Abstract

Neural network architectures for case adaptation in case-based reasoning (CBR) have received considerable attention. However, architectural gaps and general questions remain. First, existing architectures focus on adaptation of numeric attributes alone. Second, some proposed neural network adaptation architectures operate directly on pairs of cases, so could be performing direct prediction instead of adaptation. Third, it is unclear how the effectiveness of CBR systems with neural network components compares to that of networks alone. This paper addresses these questions. It extends a neural network-based case difference heuristic (NN-CDH) approach to handle both numeric and nominal attributes, in an architecture that applies to both regression and classification domains. The network predicts solution difference based on problem difference, ensuring that it learns adaptations. The paper presents experiments for both classification and regression tasks that compare performance of a neural network to a baseline CBR system and CBR variants with different retrieval schemes and adaptation schemes, on both real data and controlled artificial data sets. In these tests, CBR with the extended NN-CDH generally performs comparably to the baseline neural network, and NN-CDH consistently improves the results from naive retrieval but may worsen the results of network-based retrieval.