iEnsemble2: Committee machine model-based on heuristically-accelerated multiagent reinforcement learning

Abstract

Machine committees, as the name implies, are the union of more than one machine of learning in generating a solution to a given problem. During this process, several decisions must be taken to seek the generalization of the model and also to have the coordination to find a final solution at least satisfactory to the problem. At this point, agent theory plays a fundamental role, as it allows the agent’s autonomous decision-making, based on their experiences, as well as providing mechanisms to scale and distribute processing. Reinforcement learning is based on the existence of an external critic to the environment, which evaluates the action defined, but without explicitly indicating the correct action to be taken, in this way, allowing the training of agents in a gradual way and assisting in learning. This learning process can be accelerated by making use of heuristics on the problem domain. In this way, this article proposes a machine committee model based on multiagents system and learning by multi-person reinforcement accelerated by heuristics, describing the experiments performed and the results obtained.