Hybrid Deletion Policies for Case Base Maintenance

Case memory maintenance in a Case-Based Reasoning system is important for two main reasons: (1) to control the case memory size; (2) to reduce irrelevant and redundant instances that may produce inconsistencies in the Case-Based Reasoning system. In this paper we present two approaches based on deletion policies to the maintenance of case memories. The foundations of both approaches are the Rough Sets Theory, but each one applies a different policy to delete or maintain cases. The main purpose of these methods is to maintain the competence of the system and reduce, as much as possible, the size of the case memory. Experiments using different domains, most of them from the UCI repository, show that the reduction techniques maintain the competence obtained by the original case memory. The results obtained are compared with those obtained using well-known reduction techniques. Introduction and Motivation Case-Based Reasoning (CBR) systems solve problems by reusing the solutions to similar problems stored as cases in a case memory (Riesbeck & Schank 1989) (also known as case-base). However, these systems are sensitive to the cases present in the case memory and often its good competence depends on the significance of the cases stored. The aim of this paper is twofold: (1) to remove noisy cases and (2) to achieve a good generalisation accuracy. This paper presents two hybrid deletion techniques based on Rough Sets Theory. In a previous paper, we presented two reduction techniques based on these measures (Salamó & Golobardes 2001). This paper continues the initial approaches presented in the previous one, defining a competence model based on Rough sets and presenting new hybrid approaches to improve the weak points. The conclusion of the previous work was that the proposed reduction techniques were complementary, so hybrid methods will achieve a higher reduction and better competence case memories. Thus, in this paper, we present two hybrid approaches: Accuracy-Classification Case Memory (ACCM) and Negative Accuracy-Classification Case Memory (NACCM). Both reduction techniques have been introduced into our CaseBased Classifier System called BASTIAN. Copyright c © 2003, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. The paper is organized as follows. Next section introduces related work. In the next section, we explain the foundations of Rough Sets Theory used in our reduction techniques. After this section we detail the proposed reduction techniques based on deletion policies, continuing in next section describing the testbed of the experiments and the results obtained. Finally, in the last section, we present the conclusions and further work.