MASDES-DWMV: Model for Dynamic Ensemble Selection Based on Multiagent System and Dynamic Weighted Majority Voting

Abstract

One of the main challenges of machine learning algorithms is to maximize the result and generalization. Thus, the committee machines, i.e., the combination of more than one learning machine (an approach also called in the literature by ensemble), together with agent theory, become a promising alternative in this challenge. In this sense, this research proposes a dynamic selection model of machine committees for classification problems based on the multi-agent system and the weighted majority voting dynamic. This model has an agent-based architecture, with its roles and behaviors modeled and described throughout the life cycle of the multiagent system. The steps of generalization, selection, combination, and decision are performed through the behavior and interaction of agents with the environment, in the exercise of their respective roles. In order to validate the model, experiments were performed on 20 datasets from four repositories and compared with seven state-of-the-art dynamic committee selection models. At the end, the results of these experiments are presented, compared and analyzed, in which the proposed model obtained considerable gains in relation to the other models.