Zero-Shot Reinforcement Learning from Low Quality Data
This addresses the challenge of applying zero-shot RL to real-world problems with limited data, though it is incremental as it builds on existing conservatism techniques.
The paper tackles the problem of zero-shot reinforcement learning degrading with small, homogeneous datasets by proposing conservative algorithms, which outperform non-conservative methods on low-quality data and match or exceed baselines that see tasks during training.
Zero-shot reinforcement learning (RL) promises to provide agents that can perform any task in an environment after an offline, reward-free pre-training phase. Methods leveraging successor measures and successor features have shown strong performance in this setting, but require access to large heterogenous datasets for pre-training which cannot be expected for most real problems. Here, we explore how the performance of zero-shot RL methods degrades when trained on small homogeneous datasets, and propose fixes inspired by conservatism, a well-established feature of performant single-task offline RL algorithms. We evaluate our proposals across various datasets, domains and tasks, and show that conservative zero-shot RL algorithms outperform their non-conservative counterparts on low quality datasets, and perform no worse on high quality datasets. Somewhat surprisingly, our proposals also outperform baselines that get to see the task during training. Our code is available via https://enjeeneer.io/projects/zero-shot-rl/ .