Tianwei Ni

The Three Regimes of Offline-to-Online Reinforcement Learning

Li Li

Tianwei Ni

Yihao Sun

Pierre-Luc Bacon

Offline-to-online reinforcement learning (RL) has emerged as a practical paradigm that leverages offline datasets for pretraining and online… (see more) interactions for fine-tuning. However, its empirical behavior is highly inconsistent: design choices of online-fine tuning that work well in one setting can fail completely in another. We propose a stability--plasticity principle that can explain this inconsistency: we should preserve the knowledge of pretrained policy or offline dataset during online fine-tuning, whichever is better, while maintaining sufficient plasticity. This perspective identifies three regimes of online fine-tuning, each requiring distinct stability properties. We validate this framework through a large-scale empirical study, finding that the results strongly align with its predictions in 45 of 63 cases. This work provides a principled framework for guiding design choices in offline-to-online RL based on the relative performance of the offline dataset and the pretrained policy.

2025-10-01

ArXiv (preprint)

The Three Regimes of Offline-to-Online Reinforcement Learning

Li Li

Tianwei Ni

Yihao Sun

Pierre-Luc Bacon

Offline-to-online reinforcement learning (RL) has emerged as a practical paradigm that leverages offline datasets for pretraining and online… (see more) interactions for fine-tuning. However, its empirical behavior is highly inconsistent: design choices of online-fine tuning that work well in one setting can fail completely in another. We propose a stability--plasticity principle that can explain this inconsistency: we should preserve the knowledge of pretrained policy or offline dataset during online fine-tuning, whichever is better, while maintaining sufficient plasticity. This perspective identifies three regimes of online fine-tuning, each requiring distinct stability properties. We validate this framework through a large-scale empirical study, finding that the results strongly align with its predictions in 45 of 63 cases. This work provides a principled framework for guiding design choices in offline-to-online RL based on the relative performance of the offline dataset and the pretrained policy.

2025-10-01

ArXiv (preprint)

Understanding Behavioral Metric Learning: A Large-Scale Study on Distracting Reinforcement Learning Environments

Ziyan Luo

2025-05-09

rl-conference.cc/RLC/2025/Conference (accepted)

Understanding the Effectiveness of Learning Behavioral Metrics in Deep Reinforcement Learning

Ziyan Luo

A key approach to state abstraction is approximating behavioral metrics (notably, bisimulation metrics) in the observation space, and embed … (see more)these learned distances in the representation space. While promising for robustness to task-irrelevant noise shown in prior work, accurately estimating these metrics remains challenging, requiring various design choices that create gaps between theory and practice. Prior evaluations focus mainly on final returns, leaving the quality of learned metrics and the source of performance gains unclear. To systematically assess how metric learning works in deep RL, we evaluate five recent approaches. We unify them under isometric embedding, identify key design choices, and benchmark them with baselines across 20 state-based and 14 pixel-based tasks, spanning 250+ configurations with diverse noise settings. Beyond final returns, we introduce the denoising factor to quantify the encoder’s ability to filter distractions. To further isolate the effect of metric learning, we propose an isolated metric estimation setting, where the encoder is influenced solely by the metric loss. Our results show that metric learning improves return and denoising only marginally, as its benefits fade when key design choices, such as layer normalization and self-prediction loss, are incorporated into the baseline. We also find that commonly used benchmarks (e.g., grayscale videos, varying state-based Gaussian noise dimensions) add little difficulty, while Gaussian noise with random projection and pixel-based Gaussian noise remain challenging even for the best methods. Finally, we release an open-source, modular codebase to improve reproducibility and support future research on metric learning in deep RL.

2025-05-09

rl-conference.cc/RLC/2025/Conference (accepted)

Understanding the Effectiveness of Learning Behavioral Metrics in Deep Reinforcement Learning

Ziyan Luo

A key approach to state abstraction is approximating behavioral metrics (notably, bisimulation metrics) in the observation space, and embed … (see more)these learned distances in the representation space. While promising for robustness to task-irrelevant noise shown in prior work, accurately estimating these metrics remains challenging, requiring various design choices that create gaps between theory and practice. Prior evaluations focus mainly on final returns, leaving the quality of learned metrics and the source of performance gains unclear. To systematically assess how metric learning works in deep RL, we evaluate five recent approaches. We unify them under isometric embedding, identify key design choices, and benchmark them with baselines across 20 state-based and 14 pixel-based tasks, spanning 250+ configurations with diverse noise settings. Beyond final returns, we introduce the denoising factor to quantify the encoder’s ability to filter distractions. To further isolate the effect of metric learning, we propose an isolated metric estimation setting, where the encoder is influenced solely by the metric loss. Our results show that metric learning improves return and denoising only marginally, as its benefits fade when key design choices, such as layer normalization and self-prediction loss, are incorporated into the baseline. We also find that commonly used benchmarks (e.g., grayscale videos, varying state-based Gaussian noise dimensions) add little difficulty, while Gaussian noise with random projection and pixel-based Gaussian noise remain challenging even for the best methods. Finally, we release an open-source, modular codebase to improve reproducibility and support future research on metric learning in deep RL.

2025-05-09

rl-conference.cc/RLC/2025/Conference (published)

Do Transformer World Models Give Better Policy Gradients?

2024-07-08

Proceedings of the 41st International Conference on Machine Learning (published)

Bridging State and History Representations: Understanding Self-Predictive RL

Benjamin Eysenbach

Representations are at the core of all deep reinforcement learning (RL) methods for both Markov decision processes (MDPs) and partially obse… (see more)rvable Markov decision processes (POMDPs). Many representation learning methods and theoretical frameworks have been developed to understand what constitutes an effective representation. However, the relationships between these methods and the shared properties among them remain unclear. In this paper, we show that many of these seemingly distinct methods and frameworks for state and history abstractions are, in fact, based on a common idea of self-predictive abstraction. Furthermore, we provide theoretical insights into the widely adopted objectives and optimization, such as the stop-gradient technique, in learning self-predictive representations. These findings together yield a minimalist algorithm to learn self-predictive representations for states and histories. We validate our theories by applying our algorithm to standard MDPs, MDPs with distractors, and POMDPs with sparse rewards. These findings culminate in a set of preliminary guidelines for RL practitioners.

2024-01-17

ArXiv (preprint)

Bridging State and History Representations: Understanding Self-Predictive RL

Benjamin Eysenbach

Representations are at the core of all deep reinforcement learning (RL) methods for both Markov decision processes (MDPs) and partially obse… (see more)rvable Markov decision processes (POMDPs). Many representation learning methods and theoretical frameworks have been developed to understand what constitutes an effective representation. However, the relationships between these methods and the shared properties among them remain unclear. In this paper, we show that many of these seemingly distinct methods and frameworks for state and history abstractions are, in fact, based on a common idea of self-predictive abstraction. Furthermore, we provide theoretical insights into the widely adopted objectives and optimization, such as the stop-gradient technique, in learning self-predictive representations. These findings together yield a minimalist algorithm to learn self-predictive representations for states and histories. We validate our theories by applying our algorithm to standard MDPs, MDPs with distractors, and POMDPs with sparse rewards. These findings culminate in a set of preliminary guidelines for RL practitioners.

2024-01-17

ArXiv (preprint)

Bridging State and History Representations: Understanding Self-Predictive RL

Benjamin Eysenbach

Representations are at the core of all deep reinforcement learning (RL) methods for both Markov decision processes (MDPs) and partially obse… (see more)rvable Markov decision processes (POMDPs). Many representation learning methods and theoretical frameworks have been developed to understand what constitutes an effective representation. However, the relationships between these methods and the shared properties among them remain unclear. In this paper, we show that many of these seemingly distinct methods and frameworks for state and history abstractions are, in fact, based on a common idea of self-predictive abstraction. Furthermore, we provide theoretical insights into the widely adopted objectives and optimization, such as the stop-gradient technique, in learning self-predictive representations. These findings together yield a minimalist algorithm to learn self-predictive representations for states and histories. We validate our theories by applying our algorithm to standard MDPs, MDPs with distractors, and POMDPs with sparse rewards. These findings culminate in a set of preliminary guidelines for RL practitioners.

2024-01-17

ArXiv (preprint)

Bridging State and History Representations: Understanding Self-Predictive RL

Benjamin Eysenbach

Representations are at the core of all deep reinforcement learning (RL) methods for both Markov decision processes (MDPs) and partially obse… (see more)rvable Markov decision processes (POMDPs). Many representation learning methods and theoretical frameworks have been developed to understand what constitutes an effective representation. However, the relationships between these methods and the shared properties among them remain unclear. In this paper, we show that many of these seemingly distinct methods and frameworks for state and history abstractions are, in fact, based on a common idea of self-predictive abstraction. Furthermore, we provide theoretical insights into the widely adopted objectives and optimization, such as the stop-gradient technique, in learning self-predictive representations. These findings together yield a minimalist algorithm to learn self-predictive representations for states and histories. We validate our theories by applying our algorithm to standard MDPs, MDPs with distractors, and POMDPs with sparse rewards. These findings culminate in a set of preliminary guidelines for RL practitioners.

2024-01-16

ICLR.cc/2024/Conference (poster)

When Do Transformers Shine in RL? Decoupling Memory from Credit Assignment

Tianwei Ni

Michel Ma

Benjamin Eysenbach

Pierre-Luc Bacon

Reinforcement learning (RL) algorithms face two distinct challenges: learning effective representations of past and present observations, an… (see more)d determining how actions influence future returns. Both challenges involve modeling long-term dependencies. The Transformer architecture has been very successful to solve problems that involve long-term dependencies, including in the RL domain. However, the underlying reason for the strong performance of Transformer-based RL methods remains unclear: is it because they learn effective memory, or because they perform effective credit assignment? After introducing formal definitions of memory length and credit assignment length, we design simple configurable tasks to measure these distinct quantities. Our empirical results reveal that Transformers can enhance the memory capability of RL algorithms, scaling up to tasks that require memorizing observations