Multi-agent navigation based on deep reinforcement learning and traditional pathfinding algorithm

We develop a new framework for multi-agent collision avoidance problem. The framework combined traditional pathfinding algorithm and reinforcement learning. In our approach, the agents learn whether to be navigated or to take simple actions to avoid their partners via a deep neural network trained by reinforcement learning at each time step. This framework makes it possible for agents to arrive terminal points in abstract new scenarios. In our experiments, we use Unity3D and Tensorflow to build the model and environment for our scenarios. We analyze the results and modify the parameters to approach a well-behaved strategy for our agents. Our strategy could be attached in different environments under different cases, especially when the scale is large.