深層強化学習による施工機械の知能化

抄録

In recent years, deep reinforcement learning has better control performance than human in some Atari 2600 games. But it rarely has any practical examples with this method in the real world. In this paper, we propose to use deep reinforcement learning in construction domain. We cooperate with JAXA and we are in a project to make a human base in the moon. It is hard to send a worker to operate the leveling machine. So, our aim is using deep reinforcement learning to make the leveling machine can level the ground autonomously in various situations. We simulate the leveling action in a simulator and evaluate the method. Also, deep reinforcement learning comes with a prohibitive computational cost to finish the learning in some learning environment. We introduce a new method, using dropout in fully-connected layer leading to more efficient learning.