A framework for the design and development of diagnostic case-based reasoning systems

Abstract

<p>A Case-Based Reasoning System (CBR) supports reasoning from experience to solve decision problems, critique solutions, and explain anomalous situations. The issues encountered in the design and development of Case Based Reasoning (CBR) systems for ill-structured diagnostic decision problems include (1) the difficulty in describing a new case, (2) the effectiveness of retrieval of appropriate previous cases, and (3) the effectiveness of explanation and adaptation of retrieved previous cases. To this end, we proposed a framework of five methodologies that includes a front-end adaptive interface agent to assist a decision maker (DM) describe a new case, a rear-end interface agent to support the explanation of a new case, and three retrieval methodologies to enhance the effectiveness of retrieval methodologies. This framework was empirically tested by developing an application for the diagnosis and repair of complex electromechanical machinery. Ten expert troubleshooters from the industry participated in the empirical test. Our findings show that the framework significantly improves the existing CBR methodologies. The front-end adaptive interface agent aids a DM to describe a new case. It reduces the information load on the DM, and adapts its recommendations to the idiosyncracies of a DM. The rear-end interface agent provides task information feedback in support of the selection and adaptation of retrieved previous cases. This enables the DM to verify his/her decision processes. The framework includes two symbolic retrieval methodologies and one connectionist-symbolic retrieval methodology. The first symbolic retrieval methodology is based on a decision theoretic approach to the similarity assessment that reduces the need for production rules. It allows the adaptation of similarity to meet the information needs of a DM. The second symbolic retrieval methodology comprises the modified cosine matching function to assess the overall similarity of two cases. It overcomes some limitations of the widely used nearest-neighbour matching function by contrasting two previous cases and including the local domain knowledge that is typical in ill-structured diagnostic decision problems. However, the symbolic assessment of overall similarity is difficult to adapt to the idiosyncrasies of a DM. The adaptation is made possible by the third retrieval methodology. It is a connectionist-symbolic approach to the assessment of the overall similarity that allows the incremental learning of retrieval knowledge by interactions with a DM.</p>