A DEVS Based Methodological Framework for Reinforcement Learning Agent Training

Abstract

Reinforcement Learning has become one of the fastest growing fields of artificial intelligence due to the successful application of its techniques into several domains. In this way, the integration of intelligent agents based on Reinforcement Learning into information systems is a current reality. However, the way in which they “learn” requires a simulation model of the process that must be controlled to obtain large volumes of risk-free information. In this work, a methodological framework to support the training of Reinforcement Learning agents using DEVS is proposed. This framework provides the steps and elements required to implement RL Agents with the purpose of accelerating the agent learning and reducing training costs. Also, it allows modeling the dynamics of complex systems in a modular and hierarchical way, favoring the reuse of simulation components, since it is based on DEVS formalims fundamentals.