Abstract

Reinforcement Learning (RL) is an optimization method characterized by two interacting entities, the agent and the environment. The environment is a Markov Decision Process (MDP). The goal of RL is to learn how an agent should act to achieve a maximum cumulative reward in the long-term. In discrete-event simulation, the dynamic behavior of a system is represented in a model (DESM) that is executed via a simulator. The concept of Experimental Frame (EF) provides a structural approach to separating the DESM into the Model Under Study (MUS) and its experimental context. Here, we explore the integration of a discrete event MUS as an environment for RL using the concept of EF. After discussing the methodological framework, a case study using MATLAB/Simulink and the SimEvents blockset is considered. The case study starts with an introduction of the discrete-event MUS for which a control strategy shall be developed. The MUS is reused in three experiments using specific EFs. First, an EF for the design of a heuristic control strategy with ordinary simulation runs is presented. Then, based on the methodological approach, specifics of the EF are considered when using a self-implemented Q-agent and the RL toolbox of MATLAB/Simulink.