Research And Implementation Of Self-learning Algorithm Of RIMER Expert System

In order to handle uncertain information in decision making processes, a belief rule-base inference methodology using the evidential reasoning approach (RIMER) has been developed based on the D-S theory of evidence, decision theory and fuzzy set theory by Yang et al。In this thesis a RIMER expert system is explored on the basis of RIMER and expert system. RIMER is a white-box modeling approach and its internal structure allows the explicit representation of subjective expert knowledge. The parameters in a belief rule-base such as rule weights, the weights of antecedent attributes and belief degrees are determined by experts. But generally, experts may be difficult to determine these parameters objectively and precisely, which may hinder a RIMER expert system from imitating real systems. Yang et al developed and solved the learning model, but the results are not completely satisfactory. So it is necessary to design a new methodology to strengthen the learning ability of a RIMER expert system.On this basis, this dissertation is intended to design a new methodology by combining the gradient method with dichotomy method. First the decedent direction of the objective function of a nonlinear programming problem is found using the gradient method. Then the biggest step size is obtained according to the constraints of the nonlinear programming problem. Finally, the feasible step size which minimizes the objective function is generated.By analyzing the general structure of expert systems, a RIMER demo expert system with self-learning ability is designed on the basis of the RIMER theory and training methodology. In this paper, the total structure of the RIMER expert system is introduced, each module in the system is presented, and in particular the belief rule-base developing module, the forward reasoning module and the training module are designed and illustrated. Finally, the implementation of the RIMER system is illustrated through three training examples, which show that the new algorithm is efficient, simple, effective.