Ensemble Learning Improvement through Reinforcement Learning Idea

Ensemble learning is one of the learning methods to create a strong classifier through the integration of basic classifiers that includes the benefits of all of them. Meanwhile, weighting classifiers in the ensemble learning approach is a major challenge. This challenge arises from the fact that in ensemble learning all constructor classifiers are considered to be at the same level of distinguishing ability. While in different problem situations and especially in dynamic environments, the performance of base learners is affected by the problem space and data behavior. The solutions that have been presented in the subject literature assumed that problem space condition is permanent and static. While for each entry in real, the situation has changed and a completely dynamic environment is created. In this paper, a method based on the reinforcement learning idea is proposed to modify the weight of the base learners in the ensemble according to problem space dynamically. The proposed method is based on receiving feedback from the environment and therefore can adapt to the problem space. In the proposed method, learning automata is used to receive feedback from the environment and perform appropriate actions. Sentiment analysis has been selected as a case study to evaluate the proposed method. The diversity of data behavior in sentiment analysis is very high and it creates an environment with dynamic data behavior. The results of the evaluation on six different datasets and the ranking of different values of learning automata parameters reveal a significant difference between the efficiency of the proposed method and the ensemble learning literature.