A reinforcement learning algorithm for adjusting antecedent parameters and weights of fuzzy rules in a fuzzy classifier

This paper proposes a new fuzzy classifier based on reinforcement learning. A fuzzy rule based classification system is a special type of fuzzy modeling where its output is a discrete crisp value. The main challenging issue in designing fuzzy classifiers is constructing fuzzy rule base. Here, each fuzzy rule is considered as an agent who has to select the suitable class between candidate classes. It is considered a weight for each candidate class in each rule. These weights are adjusted using the proposed reinforcement learning algorithm. For each sample of training data, if the final result is true, the winner rule (agent) is rewarded and some other rules are punished based on the criteria which are defined in this paper. If the result is false, the winner rule is punished and the rules with high firing strength that have selected correct class are rewarded. Moreover, the input membership functions of rules are adjusted regarding the defined criteria which depend on punishment frequency of rules. The proposed approach is assessed on some UCI datasets. We compare our ideas in comparison with conventional reward and punishment scheme and multi-layer perceptron network. The experimental results show that our proposed approach outperforms both mentioned approaches in the terms of quality of classification and precision.