Multi-Agent Deep Reinforcement Learning-Based Algorithm for Fast Generalization on Routing Problems

We propose a fast generalization method for DQN-Routing, an algorithm based on multi-agent deep reinforcement learning that suffers from generalization problem when introduced to new topologies even if it was trained on a similar topology. The proposed method is based on the wisdom of crowds and allows the distributed routing algorithm, DQN-Routing, to generalize better to new topologies that were not seen before during training. The proposed method also aims to decrease the solution search time as the original DQN-Routing algorithm takes a long time to converge, and to increase the overall performance by minimizing the mean delivery time and total power consumption and the number of collisions. The experimental evaluation of our method proved that is capable to generalize to new topologies and outperform the DQN-Routing algorithm.