Diversity-augmented intrinsic motivation for deep reinforcement learning

In many real-world problems, reward signals received by agents are delayed or sparse, which makes it challenging to train a reinforcement learning (RL) agent. An intrinsic signal can help an agent explore such environments in the quest for novel states. this work, we propose general end-to-end diversity-augmented motivation deep encourages new states and automatically provides denser rewards. Specifically, measure diversity of adjacent under model state sequences based on determinantal point process (DPP); is coupled with straight-through gradient estimator enable differentiability. The proposed approach comprehensively evaluated MuJoCo Arcade Learning Environments (Atari SuperMarioBros). experiments show that derived from DPP accelerates early stages training Atari games SuperMarioBros. MuJoCo, improves prior techniques tasks using standard setting, achieves state-of-the-art performance 12 out 15 containing