Reinforcement Learning for Layout Planning – Automated Pathway Generation for Arbitrary Factory Layouts

As an intermediary step to enable solving the facility layout problem with reinforcement learning (RL), this paper presents an algorithm to automatically derive a path network from a given factory layout. It is represented as a graph with additional measures like pathwidth, segment length and segment orientation annotated as edge properties. The information can be utilized in subsequent processing steps to allow programmatic evaluation of the quality of generated layouts to aid in reward calculation for reinforcement learning. This is necessary since the material flow as a widely used evaluation criterion for algorithms solving the layout problem is not expressing the factory design goals clearly enough for a RL agent to learn. Material flow is typically computed based on transport costs and transport distance, which is abstracted as Euclidean distance between centre points of workstations. Using the path network derived by the presented algorithm enables a calculation of transport distances closer to the true distances travelled in the factory, thus enabling a more accurate reward signal.