Incremental knowledge acquisition for case-based reasoning
Abdus Salam Khan · UNSWorks (UNSW Sydney) · 2022
Case-Based Reasoning (CBR) is an appealing technique for developing intelligent systems. Besides its psychological plausibility and a substantial body of research during recent years, building a good CBR system remains still a difficult task. The main problems remaining are the development of suitable case retrieval and adaptation mechanisms for CBR. The major issues are how and when to capture the necessary knowledge for both of the above aspects. As a contribution to knowledge this thesis proposes a new approach to address the experienced difficulties. The basic framework of Ripple Down Rules (RDR) is extended to allow the incremental development of a knowledge base for each of the two functions: case retrieval and case adaptation, during the use of the system while solving actual problems. The proposed approach allows an expert-user to provide explanations of why, for a given problem, certain actions should be taken. Incrementally knowledge is acquired from the expert-user in which the expert refines a rule which performs unsatisfactorily for a current given problem. The approach facilitates both, the rule acquisition as well as its validation. As a result the knowledge maintenance task of a knowledge engineer is overcome. This approach is effective with respect to both, the development of highly tailored and complex retrieval and adaptation functions for CBR as well as the provision of an intuitive and feasible approach for the expert. The approach has been implemented in a CBR system named MIKAS (Menu Construction using Incremental Knowledge Acquisition Systems) for the design of menus (diets) according to dietary requirements. The experimental evidence indicates the suitability of the approach to address the retrieval and adaptation problems of the menu construction domain. The experimental evidence also indicates that the difficulties of developing retrieval and adaptation functions for CBR can be effectively overcome by the proposed new approach. It is expected that the approach is likely to be useful in other problem solving domains where expert intervention is required to modify a solution.