Sequence-level Intrinsic Exploration Model for Partially Observable Domains

Training reinforcement learning policies in partially observable domains with sparse reward signal is an important and open problem for the research community. In this paper, we introduce a new sequence-level intrinsic novelty model to tackle the challenge of training reinforcement learning policies in sparse rewarded partially observable domains. First, we propose a new reasoning paradigm to infer the novelty for the partially observable states, which is built upon forward dynamics prediction. Different from conventional approaches that perform self-prediction or one-step forward prediction, our proposed approach engages open-loop multi-step prediction, which enables the difficulty of novelty prediction to flexibly scale and thus results in high-quality novelty scores. Second, we propose a novel dual-LSTM architecture to facilitate the sequence-level reasoning over the partially observable state space. Our proposed architecture efficiently synthesizes information from an observation sequence and an action sequence to derive meaningful latent representations for inferring the novelty for states. To evaluate the efficiency of our proposed approach, we conduct extensive experiments on several challenging 3D navigation tasks from ViZDoom and DeepMind Lab. We also present results on two hard-exploration domains from Atari 2600 series in Appendix to demonstrate our proposed approach could generalize beyond partially observable navigation tasks. Overall, the experiment results reveal that our proposed intrinsic novelty model could outperform several state-of-the-art curiosity baselines with considerable significance in the testified domains.