Multiagent Reinforcement Learning in Escape Scenario
Donghun Lee, Seonghyun Kim, Young-Sung Son · 2018
Reinforcement learning let agents perform repeated actions in given environments and identify how to behave in the given environments. In this paper, we showed possibilities of solutions for real-world scenarios on multiagent reinforcement learning environment. Our experiment showed an approach to investigate a way for agents to escape in an escape scenario using multiagent reinforcement learning. Two types of agents are trained. We trained Escaper Agent to find an optimal exit which increases a total number of escaped agents and decreases congestion and Exit Agent to find optimal locations for exits. Thoroughly setting reward function could simulate real-world problems and get insights on how to break down challenging questions.