Revisiting Bellman Errors for Offline Model Selection
This work addresses a crucial challenge for applying offline RL in real-world settings by improving policy selection from logged data, though it is incremental as it builds on existing Bellman error approaches.
The paper tackles the problem of offline model selection in reinforcement learning by analyzing why previous Bellman error-based methods underperformed and developing a new, more accurate estimator. Their estimator achieves strong performance on diverse discrete control tasks, including Atari games.
Offline model selection (OMS), that is, choosing the best policy from a set of many policies given only logged data, is crucial for applying offline RL in real-world settings. One idea that has been extensively explored is to select policies based on the mean squared Bellman error (MSBE) of the associated Q-functions. However, previous work has struggled to obtain adequate OMS performance with Bellman errors, leading many researchers to abandon the idea. To this end, we elucidate why previous work has seen pessimistic results with Bellman errors and identify conditions under which OMS algorithms based on Bellman errors will perform well. Moreover, we develop a new estimator of the MSBE that is more accurate than prior methods. Our estimator obtains impressive OMS performance on diverse discrete control tasks, including Atari games.