Dissecting Deep RL with High Update Ratios: Combatting Value Divergence
This work addresses a fundamental challenge in deep RL for researchers and practitioners, offering a simple solution to improve learning efficiency in data-scarce settings, though it is incremental as it builds on prior studies of primacy bias.
The paper tackled the problem of deep reinforcement learning failing under high update-to-data ratios by identifying value function divergence as the core issue, and demonstrated that a simple unit-ball normalization enables effective learning, achieving strong performance on challenging dm_control tasks competitive with model-based approaches.
We show that deep reinforcement learning algorithms can retain their ability to learn without resetting network parameters in settings where the number of gradient updates greatly exceeds the number of environment samples by combatting value function divergence. Under large update-to-data ratios, a recent study by Nikishin et al. (2022) suggested the emergence of a primacy bias, in which agents overfit early interactions and downplay later experience, impairing their ability to learn. In this work, we investigate the phenomena leading to the primacy bias. We inspect the early stages of training that were conjectured to cause the failure to learn and find that one fundamental challenge is a long-standing acquaintance: value function divergence. Overinflated Q-values are found not only on out-of-distribution but also in-distribution data and can be linked to overestimation on unseen action prediction propelled by optimizer momentum. We employ a simple unit-ball normalization that enables learning under large update ratios, show its efficacy on the widely used dm_control suite, and obtain strong performance on the challenging dog tasks, competitive with model-based approaches. Our results question, in parts, the prior explanation for sub-optimal learning due to overfitting early data.