[4G2-GS-7] Agents: Agents and machine learning

[4G2-GS-7-01] Analysis of Q-Network structure in deep Q learning to optimize elevator control

〇Wataru Yoshida¹, Naoki Mori¹ (1. Osaka Prefecture University)

Keywords:Elevator, Deep Q-Learning, Reinforcement learning

In recent years, deep reinforcement learning has made remarkable progress. In this study, I focused on an elevator control problem, one of the important reinforcement learning tasks. In this problem, in a building with multiple elevators, it is necessary to optimize the behavior of the car so that the waiting time for passengers is reduced. Therefore, approaches using deep reinforcement learning for optimization of the controller have recently attracted attention. In this study, I aimed to improve passenger's transportation efficiency on a time average by modifying the controller in the previous study, which is implemented using Deep Q-Network (DQN), one of the deep reinforcement learning. So I propose two Q-Network structures, which have the role of determining Q score in DQN. From the experimental results, it is confirmed that both of the proposed Q-Network structures function more effectively than the conventional method in a specific time zone.