Performance Comparison of Single-Agent and Multi-Agent Reinforcement Learning for Factory Layout Planning

The result of the factory layout planning process has a direct impact on the operating costs of the future manufacturing systems. As manual layout planning is a time-consuming process, automated approaches have been developed to optimize factory layouts under varying boundary conditions and optimization objectives. Recent developments show that single-agent reinforcement learning (RL) approaches are capable of optimizing factory layouts. Extending the existing single-agent approache to a multi-agent RL approach offers performance-related potential that has not been investigated so far. The use of a multi-agent RL approach which can distribute the task of layout optimization can lead to improved layout variants. However, building a multi-agent approach also involves several challenges, such as partial observability and competition among agents. It is not clear which approach is more suitable for factory layout planning and under what conditions one approach can be chosen over the other. This paper addresses this research gap by comparing a single-agent and a multi-agent RL approach for factory layout planning. The performance of both approaches is compared based on an application scenario of 30 functional units, using the transportation intensity as the optimization criterion. The results show that the single-agent RL approach outperforms the multi-agent RL approach slightly. One conclusion is that the advantages of using a multi-agent RL approach can not overcome the corresponding challenges.