An Assessment of Different Reinforcement Learning Methods for Creating a Decision Support System Based on the Petri Net Model

Abstract

Having an optimal decision support system (DSS) is beneficial for decreasing the lifecycle costs and increasing the lifespan of any system. Reinforcement learning can serve as a good tool for evaluating the benefit of taking decisions and controlling actions to reach an optimal policy. Besides, Petri nets are adequate for modeling complex systems while taking minimal assumptions due to their capabilities of simulating heterogeneous information, parallel operations, and synchronization. In addition to having the benefit of providing a visual interpretation based on a well-formulated mathematical model. Combining the two concepts of Reinforcement learning and Petri net is fruitful for creating an intelligent DSS that can decide on the right actions with minimal human intervention. RL is powerful in doing so because it considers the final goal without getting into the rules of the problem, relations between events, or any other complications. It simplifies the problem of setting rewards and finding an optimal policy for increasing them. This study proposes a methodology for creating a DSS for modeling and optimizing maintenance strategy by merging RL and PN models. It also compares the advantages and disadvantages of using each of the Monte Carlo and Q-learning RL techniques and investigates the power of the n-step bootstrapping method which is a generalization of the latter methods that considers both of them at the same time.