Curiosity-Driven Variational Autoencoder for Deep Q Network

In recent years, deep reinforcement learning (DRL) has achieved tremendous success in high-dimensional and large-scale space control and sequential decision-making tasks. However, the current model-free DRL methods suffer from low sample efficiency, which is a bottleneck that limits their performance. To alleviate this problem, some researchers used the generative model for modeling the environment. But the generative model may become inaccurate or even collapse if the state has not been sufficiently explored. In this paper, we introduce a model called Curiosity-driven Variational Autoencoder (CVAE), which combines variational autoencoder and curiosity-driven exploration. During the training process, the CVAE model can improve sample efficiency while curiosity-driven exploration can make sufficient exploration in a complex environment. Then, a CVAE-based algorithm is proposed, namely DQN-CVAE, that scales CVAE to higher dimensional environments. Finally, the performance of our algorithm is evaluated through several Atari 2600 games, and the experimental results show that the DQN-CVAE achieves better performance in terms of average reward per episode on these games.