IGN : Implicit Generative Networks
This work addresses distributional reinforcement learning for Atari games, offering incremental improvements in performance and risk-sensitive policy training.
The paper tackles the problem of approximating full quantile value distributions for state-action returns in reinforcement learning by integrating GAN components with quantile regression, achieving improved performance on 57 Atari 2600 games and demonstrating state-of-the-art training for risk-sensitive policies.
In this work, we build recent advances in distributional reinforcement learning to give a state-of-art distributional variant of the model based on the IQN. We achieve this by using the GAN model's generator and discriminator function with the quantile regression to approximate the full quantile value for the state-action return distribution. We demonstrate improved performance on our baseline dataset - 57 Atari 2600 games in the ALE. Also, we use our algorithm to show the state-of-art training performance of risk-sensitive policies in Atari games with the policy optimization and evaluation.