Deep Adversarial Reinforcement Learning With Noise Compensation by Autoencoder

We present a new adversarial learning method for deep reinforcement learning (DRL). Based on this method, robust internal representation in a deep Q-network (DQN) was introduced by applying adversarial noise to disturb the DQN policy; however, it was compensated for by the autoencoder network. In particular, we proposed the use of a new type of adversarial noise: it encourages the policy to choose the worst action leading to the worst outcome at each state. When the proposed method, called deep Q-W-network regularized with an autoencoder (DQWAE), was applied to seven different games in an Atari 2600, the results were convincing. DQWAE exhibited greater robustness against the random/adversarial noise added to the input and accelerated the learning process more than the baseline DQN. When applied to a realistic automatic driving simulation, the proposed DRL method was found to be effective at rendering the acquired policy robust against random/adversarial noise.