Multiagent algorithm for fuzzy rule bases design for classification problem

In this article a multiagent approach for organizing an ensemble of optimizers, based on the meetings between algorithms is presented. During the optimization process the agents exchange with best solutions and better algorithms receive more resources in the form of meetings. Among agents the six genetic algorithms with different operators and three particle swarm optimizers have been selected. The proposed approach of ensemble-based optimization problems solving is applied to the problem of designing a fuzzy rule base. The fuzzy rule base consisted of a fixed number of rules, for every variable and every rule the membership function was defined with two sigmoidal functions. The encoded parameters were the points where sigmoids reached 0 and 1, so that the problem of designing a fuzzy rule base reduced to a real-valued optimization problem. The number of real-valued parameters depended on the dimension of the classification problem. The effectiveness of the algorithm was compared to the self-configured genetic algorithm, solving the same problem of designing a fuzzy rule base. The classification quality was estimated using the accuracy values; the sample was split with 70 and 30 ratio. As classification problems, six problems have been selected from KEEL and UCI repositories, including credit scoring problems, medical diagnostics problems, banknote recognition and seeds' forms. Two more classification methods have been selected for comparison, more precisely, support vector machines (SVM) and another fuzzy classification method, in which the term numbers were encoded. According to the testing results, it should be mentioned that the multiagent algorithm has shown the effectiveness, comparable to other method when solving complex optimization problems.