GYM_PM: A REINFORCEMENT LEARNING FRAMEWORK FOR PREDICTIVE MAINTENANCE AND SUPPLY CHAIN OPTIMISATION

Abstract

The paper provides an overview of the Python package that we have developed for Deep Reinforcement Learning (DRL) in the domains of Predictive Maintenance and Supply Chain Optimisation. The package consists of 4 different environments in the categories of Rolling Stock and Manufacturing. We went on to benchmark algorithms such as PPO and IMPALA against heuristic baselines to evaluate the performance of our agents. We found that on-policy algorithms tend to outperform off-policy algorithms on more complex scenarios while the opposite is true for simpler environments. Lastly, we propose further enhancements that can be made to the design of our environments.