Training environments for Reinforcement Learning Agents in Discrete Event System Specification

Abstract

Nowadays, dynamics and complexity of industrial environments have led to the need for solutions that allow capturing the interaction in real-time to make decisions about the control of the involved processes. Reinforcement Learning is a promising approach to solve sequential decision problems, where the complexity lies in the agent-environment interaction and the underlying uncertainty of the environment. This requires a simulation that reflects the process under control (environment) and its dynamics to train the agent. In this work, a novel solution is presented to train this type of agent with modeled and simulated environments using Discrete Event System Specification. The same applies to the problem of generation and administration of alternative energy, biogas produced by a digester and used by different industrial consumer profiles.