[3Xin2] Poster session 1

[3Xin2-18] Robustness Verification of Transformer-based Reinforcement Learning for Atari Games

〇Tsuyoshi Takano¹, Hiroshi Kera², Kazuhiko Kawamoto² (1.Graduate School of Science and Engineering, Chiba University, 2.Graduate School of Engineering, Chiba University)

Keywords:Reinforcement Learning, Transformer, Atari, Robustness

In this study, we examine the robustness of Transformer-based offline reinforcement learning. We train the state data for offline reinforcement learning with noise and evaluate the performance in training under noise. In the evaluation experiments, we compare the performance of four different Atari games (Breakout, Pong, Qbert, and Seaquest) in terms of scores on five evaluation tests (Clean, Gaussian, Shot, Impulse, and Speckle). The experimental results showed that the Atari game scores were lower on all noise evaluation tests for normal training (clean). The results showed that the Atari game score tended to improve when data augmentation training with a noise system was introduced during training. This result indicates the vulnerability of Atari games to noise evaluation tests and the robustness of Atari games improved by data augmentation training.