Deep Reinforcement Learning for Multi-agent Simulation using a partial floor field cutout

The purpose of this study is to easily create agent behavior for disaster evacuation simulation. Multi-agent simulation is commonly used in disaster evacuation simulations, but it is difficult to program agents. Floor field models and reinforcement learning have been proposed as a way of solving this problem. However, there are issues with unnatural stagnation or learning times. In this paper, we report the reinforcement learning method for an evacuation simulation agent, using the local environment. As a result of the experiment, it was confirmed that efficient movement of the agent can be achieved in a short learning time.