Showing Your Offline Reinforcement Learning Work: Online Evaluation Budget Matters
This work addresses the problem of unreliable comparisons in offline RL for researchers and practitioners, highlighting an incremental improvement in evaluation methodology.
The paper argues that online evaluation budgets are crucial for reliable comparisons of deep offline reinforcement learning algorithms, demonstrating that algorithm preferences vary with budget across domains like Robotics, Finance, and Energy Management. It proposes using Expected Validation Performance from NLP to estimate performance under different budgets without extra computation, showing that Behavioral Cloning often outperforms offline RL under limited budgets.
In this work, we argue for the importance of an online evaluation budget for a reliable comparison of deep offline RL algorithms. First, we delineate that the online evaluation budget is problem-dependent, where some problems allow for less but others for more. And second, we demonstrate that the preference between algorithms is budget-dependent across a diverse range of decision-making domains such as Robotics, Finance, and Energy Management. Following the points above, we suggest reporting the performance of deep offline RL algorithms under varying online evaluation budgets. To facilitate this, we propose to use a reporting tool from the NLP field, Expected Validation Performance. This technique makes it possible to reliably estimate expected maximum performance under different budgets while not requiring any additional computation beyond hyperparameter search. By employing this tool, we also show that Behavioral Cloning is often more favorable to offline RL algorithms when working within a limited budget.