Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluation-Time Policy Switching for Offline Reinforcement Learning

Natinael Solomon Neggatu, Jeremie Houssineau, Giovanni Montana

TL;DR

This paper tackles offline reinforcement learning by addressing the core challenge of safely generalising from fixed datasets to unseen situations. It introduces an evaluation-time policy switching framework that combines a pure offline RL policy with a behavioural cloning policy, guided by epistemic uncertainty from an ensemble of critics and a dataset-derived measure of aleatoric uncertainty. The switching mechanism defines a pessimistic RL value and a data-driven penalty to select between $a_{RL}$ and $a_{BC}$, enabling robust performance across diverse datasets without retraining, and it naturally extends to online fine-tuning via a REDQ-inspired update strategy. Empirically, the approach yields strong offline performance and smooth online adaptation on D4RL MuJoCo and AntMaze benchmarks, with ablations confirming the benefit of incorporating dataset diversity into the uncertainty penalty. Overall, the method provides a flexible, low-tuning pathway to harness the strengths of both RL and BC for offline-to-online RL.

Abstract

Offline reinforcement learning (RL) looks at learning how to optimally solve tasks using a fixed dataset of interactions from the environment. Many off-policy algorithms developed for online learning struggle in the offline setting as they tend to over-estimate the behaviour of out of distributions actions. Existing offline RL algorithms adapt off-policy algorithms, employing techniques such as constraining the policy or modifying the value function to achieve good performance on individual datasets but struggle to adapt to different tasks or datasets of different qualities without tuning hyper-parameters. We introduce a policy switching technique that dynamically combines the behaviour of a pure off-policy RL agent, for improving behaviour, and a behavioural cloning (BC) agent, for staying close to the data. We achieve this by using a combination of epistemic uncertainty, quantified by our RL model, and a metric for aleatoric uncertainty extracted from the dataset. We show empirically that our policy switching technique can outperform not only the individual algorithms used in the switching process but also compete with state-of-the-art methods on numerous benchmarks. Our use of epistemic uncertainty for policy switching also allows us to naturally extend our method to the domain of offline to online fine-tuning allowing our model to adapt quickly and safely from online data, either matching or exceeding the performance of current methods that typically require additional modification or hyper-parameter fine-tuning.

Evaluation-Time Policy Switching for Offline Reinforcement Learning

TL;DR

This paper tackles offline reinforcement learning by addressing the core challenge of safely generalising from fixed datasets to unseen situations. It introduces an evaluation-time policy switching framework that combines a pure offline RL policy with a behavioural cloning policy, guided by epistemic uncertainty from an ensemble of critics and a dataset-derived measure of aleatoric uncertainty. The switching mechanism defines a pessimistic RL value and a data-driven penalty to select between and , enabling robust performance across diverse datasets without retraining, and it naturally extends to online fine-tuning via a REDQ-inspired update strategy. Empirically, the approach yields strong offline performance and smooth online adaptation on D4RL MuJoCo and AntMaze benchmarks, with ablations confirming the benefit of incorporating dataset diversity into the uncertainty penalty. Overall, the method provides a flexible, low-tuning pathway to harness the strengths of both RL and BC for offline-to-online RL.

Abstract

Offline reinforcement learning (RL) looks at learning how to optimally solve tasks using a fixed dataset of interactions from the environment. Many off-policy algorithms developed for online learning struggle in the offline setting as they tend to over-estimate the behaviour of out of distributions actions. Existing offline RL algorithms adapt off-policy algorithms, employing techniques such as constraining the policy or modifying the value function to achieve good performance on individual datasets but struggle to adapt to different tasks or datasets of different qualities without tuning hyper-parameters. We introduce a policy switching technique that dynamically combines the behaviour of a pure off-policy RL agent, for improving behaviour, and a behavioural cloning (BC) agent, for staying close to the data. We achieve this by using a combination of epistemic uncertainty, quantified by our RL model, and a metric for aleatoric uncertainty extracted from the dataset. We show empirically that our policy switching technique can outperform not only the individual algorithms used in the switching process but also compete with state-of-the-art methods on numerous benchmarks. Our use of epistemic uncertainty for policy switching also allows us to naturally extend our method to the domain of offline to online fine-tuning allowing our model to adapt quickly and safely from online data, either matching or exceeding the performance of current methods that typically require additional modification or hyper-parameter fine-tuning.

Paper Structure

This paper contains 25 sections, 6 equations, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Offline RL
  4. Offline to Online Fine-tuning
  5. Methodology
  6. Preliminaries
  7. Policy switching
  8. Batch constrained MDP
  9. Uncertainty quantification
  10. Epistemic uncertainty
  11. Aleatoric uncertainty
  12. Action switching mechanism
  13. Offline to online fine-tuning
  14. Experimental results
  15. Offline RL tasks and datasets
  16. ...and 10 more sections

Figures (2)

  • Figure 1: The top row shows the learning curves for online fine-tuning on the MuJoCo datasets. Algorithms are evaluated every 10k steps and averaged over 5 seeds and 10 episodes per seed. Bottom row shows how standard deviation across the ensemble of critic evolves for our algorithm in contrast to TD-N (blue and yellow line y-axis LHS) as well as the average proportion of BC used across an episode for our algorithm (red line y-axis RHS).
  • Figure 2: The top row shows fine-tuning curves on the AntMaze datasets. Algorithms are evaluated every 10k steps averaged over 3 seeds and 100 episodes per seed. Bottom row shows evolution of standard deviation across the ensemble of critics and the average proportion of BC.