[2C5-GS-2] Machine learning: reinforcement learning (2)

[2C5-GS-2-04] Safe Reinforcement Learning in Continuous State Spaces based on DQN

〇Yuto Ohashi¹, Tohgoroh Matsui², Atsuko Mutoh¹, Koichi Moriyama¹, Nobuhiro Inuzuka¹ (1. Nagoya Institute of Technology, 2. Chubu University)

Keywords:AI, machine learning, reinforcement learning, safe reinforcement learning, deep learning

These days, we use robots in dangerous environments where we can't go in, such as disaster areas and space. Safety reinforcement learning learns how to avoid dangers. Umemoto et al. has proposed HDEQ of Safety Reinforcement Learning in high-dimensional continuous space. HDEQ uses VAE to decrease dimensions of observation, but this doesn't work on the problems such as robot control that we want to solve. We propose a method based on DQN to solve these problems. In addition, we confirm whether the proposed method can avoid dangers in a Safety Gym environment or not.