Position: Lifetime tuning is incompatible with continual reinforcement learning

Golnaz Mesbahi, Parham Mohammad Panahi, Olya Mastikhina, Steven Tang, Martha White, Adam White

TL;DR: Continual reinforcement learning research is being mislead by lifetime tuning.

Abstract: In continual RL we want agents capable of never-ending learning, and yet our evaluation methodologies do not reflect this. The standard practice in RL is to assume unfettered access to the deployment environment for the full lifetime of the agent. For example, agent designers select the best performing hyperparameters in Atari by testing each for 200 million frames and then reporting results on 200 million frames. In this position paper, we argue and demonstrate the pitfalls of this inappropriate empirical methodology: lifetime tuning. We provide empirical evidence to support our position by testing DQN and SAC across several of continuing and non-stationary environments with two main findings: (1) lifetime tuning does not allow us to identify algorithms that work well for continual learning---all algorithms equally succeed; (2) recently developed continual RL algorithms outperform standard non-continual algorithms when tuning is limited to a fraction of the agent's lifetime. The goal of this paper is to provide an explanation for why recent progress in continual RL has been mixed and motivate the development of empirical practices that better match the goals of continual RL.

Lay Summary: In continual reinforcement learning (RL) we want agents capable of never-ending learning, and yet our evaluation methodologies do not reflect this. The standard practice in RL is to assume unfettered access to the deployment environment for the full lifetime of the agent. For example, agent designers select the best performing hyperparameters (configurable parameters) in Atari by testing each for 200 million frames and then reporting results on 200 million frames. In this position paper, we argue and demonstrate the pitfalls of this inappropriate empirical methodology: \emph{lifetime tuning}. We provide empirical evidence to support our position by testing DQN and SAC (two popular algorithms) across several continuing and non-stationary environments with two main findings: (1) lifetime tuning does not allow us to identify algorithms that work well for continual learning---all algorithms equally succeed; (2) recently developed continual RL algorithms outperform standard non-continual algorithms when tuning is limited to a fraction of the agent's lifetime. The goal of this paper is to provide an explanation for why recent progress in continual RL has been mixed and motivate the development of empirical practices that better match the goals of continual RL.

Verify Author Names: My co-authors have confirmed that their names are spelled correctly both on OpenReview and in the camera-ready PDF. (If needed, please update ‘Preferred Name’ in OpenReview to match the PDF.)

No Additional Revisions: I understand that after the May 29 deadline, the camera-ready submission cannot be revised before the conference. I have verified with all authors that they approve of this version.

Pdf Appendices: My camera-ready PDF file contains both the main text (not exceeding the page limits) and all appendices that I wish to include. I understand that any other supplementary material (e.g., separate files previously uploaded to OpenReview) will not be visible in the PMLR proceedings.

Latest Style File: I have compiled the camera ready paper with the latest ICML2025 style files <https://media.icml.cc/Conferences/ICML2025/Styles/icml2025.zip> and the compiled PDF includes an unnumbered Impact Statement section.

Paper Verification Code: ZDg4N

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: Continual Reinforcement Learning, Lifelong Reinforcement Learning, Hyperparameter Selection

Submission Number: 428