Bootstrap State Representation Using Style Transfer for Better Generalization in Deep Reinforcement Learning

Deep Reinforcement Learning (RL) agents often overfit the training environment, leading to poor generalization performance. In this paper, we propose Thinker, a bootstrapping method remove adversarial effects of confounding features from observation in an unsupervised way, and thus, it improves RL agents’ generalization. Thinker first clusters experience trajectories into several clusters. These are then bootstrapped by applying style transfer generator, which translates one cluster’s another while maintaining content observations. The used for policy learning. has wide applicability among many settings. Experimental results reveal that leads better capability Procgen benchmark environments compared base algorithms data augmentation techniques.