Förstärkningsinlärning som designer av anläggningsmaskinens motorutrymme

Designing an engine compartment of a construction machine in CAD is today mainly a manual and time-consuming process and there are few automated tools available within design software programs. Therefore, it is investigated how reinforcement learning (RL) could be applied to design the complex engine compartment, with a focus on the placement of components and the routing of wires between them. A practical example is presented where the policy gradient method PPO was applied to learning an agent that would place five different objects in a discrete two-dimensional area bounded by several conditions. The results showed success to some extent, however due to conflicts among the reward functions the agent had difficulty achieving the objectives. Further, multi-agent RL (MARL) was investigated in combination with deep learning, curriculum learning, feudal RL and transfer-learning with the hope of finding an approach that leads to a harmonious merging of component placement and joint drawing.