Reinforcement Learning from Simulated Environments: An Encoder Decoder Framework

Reinforcement learning (RL) is used for sequential decision making such as operating and maintaining manufacturing systems. In RL, the system is modeled as a Markov decision process with states, actions, rewards, and policies. A policy is learned through repeated interaction with the environment. When a RL agent cannot interact with the real system due to time and cost constraints, a simulation of the system may be used in its place. Unfortunately, most simulations are not built for the purpose of interacting with a RL agent. Simulations built to function as the environment are often structured only according to the defined state, action, and reward and can lack fidelity, detail, and accuracy. We propose a general framework for bridging the worlds of simulation and RL. This is accomplished by placing "interpreters" between the simulation and the RL agent that translate information into a form that is coherent to each entity.