Implementing an Online Scheduling Approach for Production with Multi Agent Proximal Policy Optimization (MAPPO)

The manufacturing process relies on a well-coordinated schedule that optimally incorporates all available resources to achieve maximum profit. In the case of machine breakdowns, the created schedule does not contain information on how to proceed further. Manual adjustments to the process order do not guarantee optimal utilization of the available resources, as many interconnections of the manufacturing process are not evident to a human. A reliable method is needed that can react to changing conditions on the shop floor and form well-founded decisions to mitigate negative effects. This paper presents an approach to implement Multi Agent Reinforcement Learning for online scheduling a cell-based manufacturing environment with unpredictable machine breakdowns. The developed “Multi Agent Proximal Policy Optimization”-Algorithm (MAPPO) combines already existing approaches in a novel way, by using the centralized learning and decentralized execution together with an objective function developed for OpenAIs Proximal Policy Optimization algorithm.