Dynamic Case Memory Management

The present work describes some aspects related to the utility/swamping problem in ADAPtER, a multimodal reasoning system combining Case-Based Reasoning and Model-Based Reasoning for diagnostic problem solving. A detailed set of experiments allowed us to analyse the average behavior of the system with respect to a given domain, in terms of performance of the whole system and its components. Such experiments pointed out that the increasing of the size of the case memory reduces the need for solving problem from scratch, but is the main responsible for the arising of the utility problem in ADAPtER. As a consequence, particular attention is paid to the problem of dynamically maintaining under control the growth of the case memory. We propose two learning strategies implementing a dynamic approach to case memory management. Such strategies allow the system to dynamically add or replace cases from memory, in order to keep under control both case memory size and content. Experimental testing of the above strategies suggests that their adoption can greatly mitigate the over-sizing of the case memory.