Abstract
As product systems become larger and more complex, their design space is increasing, and it is becoming more difficult to optimize them. In the past, hierarchical optimization methods have been proposed to solve this problem by decomposing the entire system into several subsystems and dividing and reducing the design space. However, they are ineffective for problems with strong dependencies among many subsystems. On the other hand, it has become clear that reinforcement learning can be used to efficiently search for optimal solutions to optimization problems such as solving small-scale games. If we try to apply reinforcement learning to large-scale optimization problems, the learning space of reinforcement learning increase, and optimization become difficult. Therefore, in this study, we used multi-agent reinforcement learning to reduce the learning space handled by reinforcement learning at a time. In this method, we propose an algorithm that reduces the learning space by allowing each agent to learn only the priority of a subsystem based on the relationship between the design variables and the evaluation values of each subsystem. The proposed method reduced the number of evaluations to less than 10% of the conventional number while maintaining the solution quality of the optimization.
