A Hybrid Case-Based Reasoning Architecture and It''s Application

• Main Authors: Chien-Chang Hsu, 許見章
• Other Authors: Cheng-Seen Ho
• Format: Others
• Language: en_US
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/22359028109768658469

博士 === 國立臺灣科技大學 === 電子工程系 === 88 === Case-based reasoning (CBR) is a problem solving method that maps problem features to potential solutions through the process of case retrieval, case selection, case adaptation, and case retaining. However, a solution may get involved in a huge number of problem features. How do we know which are relevant or significant? How do we correctly and completely identify them? How do we efficiently retrieve relevant past cases that are most worth adaptation without too much adaptation effort? How do we manipulate candidate cases in order to generate a new solution that fits a given problem? Finally, how do we maintain a case library so that it only assimilates worthy adapted cases? These are "classical" issues associated with CBR. Currently published CBR systems have proposed a variety of hybridization techniques to solve all or some of the issues. Their efforts, however, still haven’t promoted CBR into one of the main stream intelligent problem solving methods. We noticed that following shortcomings are associated with current CBR systems and need to be carefully addressed. First, most existent CBR systems pre-assume that the user can provide correct information about a given problem. It assumes that the collected problem data is correct and complete. This explains why the quality of their solutions is sometimes stuck in an unsatisfactory level. Some CBR systems do provide a better user interface. However, the interface is either inflexible or template-like for specific users. It is rather hard for such systems to obtain correct data input from different proficiency levels of users. Recognizing that users with different backgrounds tend to talk in different languages, to successfully solicit data from them needs more delicate design in the interface. Second, most existent CBR systems use a hierarchy structure to index the cases in the case library. The hierarchy structure extracts common features into a prototype that classifies cases into different groups or classes to facilitate search of similar cases. It is, however, not appropriate for a static index hierarchy, to be used in different types of problems since the significance of a feature may change in different context. Third, some CBR systems use surface features and complex similarity measurement to find the most similar case for a given problem without considering the adaptability of the case. The most similar case, however, does not guarantee to be the most adaptable. Fourth, many CBR systems use context-dependent adaptation knowledge to do case adaptation. The adaptation mechanisms are ad hoc and few of them can be re-used in other domains. Finally, very fewer of CBR systems can discover new adaptation knowledge from the newly adapted case. This thesis proposes a hybrid CBR architecture to alleviate the above issues. The architecture contains two modules, namely, an intelligent interface agent and an HCBR. The intelligent interface agent provides an adaptable and adaptive human-machine interaction pattern to solicit data from the user. Specifically, it supports different styles of interaction patterns and helping functions for the user according to his proficiency in the specific domain. It also supports input data rational checking and supports answers for less-experienced users. A domain-related common sense knowledge base and a past Q&A scenarios repository are behind the above two operations. A data analyst is defined as the major mechanism that discriminates the proficiency levels of the user; check for completeness, consistency, and rationality of input data; selects human-machine interaction styles; and manages reasoning. With these techniques, the intelligent interface agent is able to help the doctors of different levels of domain proficiency to gather correct and complete patient data. The HCBR module hybridizes CBR, fuzzy neural networks, induction, utilities-based decision theory, and knowledge-based planning technology to facilitate solutions finding. The basic mechanism is CBR that accumulates experiences as cases in the case library and proposes solutions by adapting the old cases that have successfully solved previous case. The distributed fuzzy neural network performs approximate matching to tolerate potential noise in case retrieval. The induction technology along with relevance theory is used in case selection, adaptation and learning, which helps a lot in hammering out valuable features for the target case from existent ones and in pruning unnecessary search space. Knowledge-based planning is used as a general architecture for case adaptation. It creates an adaptation plan from an adaptation tree that covers all the relevant problem features, satisfies all the relevant constraints, and contains all cases whose expected utility are over a threshold. Execution of the case adaptation plan thus can successfully propose diagnosed diseases. Moreover, the adaptation tree can support case reuse and learning of various knowledge, including case-specific association knowledge, adaptation rules, and differential rules, to enhance subsequent CBR reasoning. Hybridizing these techniques in the CBR module can effectively produce a high-quality diagnosis for a given medical consultation. We have applied the proposed techniques in the medicine domain to do medical diagnosis with multiple diseases. The experiment shows that the architecture can successfully interact with the physician with different proficiency levels in collecting complete patient data, which in turns leads to a correct diagnosis of multiple diseases. The best thing is that it also successfully discovers new medical adaptation knowledge from the relevant cases. The application of the technique in the medicine domain not only improves the quality of the diagnosed diseases but reduces the burden of a physician.