CODS Selects Optimal Data for Reusable Offline Reinforcement Learning

Ibne Farabi Shihab, Sanjeda Akter, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Md Najmus Swaqeeb, Anuj Sharma· August 11, 2026 View original

Key takeaways

  • Offline RL training can be inefficient due to redundant data.
  • CODS iteratively selects high-residual data, creating a reusable, efficient subset.
  • It significantly improves performance and reduces computational costs compared to other methods.
  • CODS is particularly beneficial for long-horizon tasks requiring rare transitions.

Who benefits

RoboticsAutonomous VehiclesGamingLogisticsPersonalized Recommendations

Summary

CODS (Critic-Guided Selector) is a new method for offline reinforcement learning that iteratively selects a high-residual, reusable subset of training data. It significantly improves performance and computational efficiency by focusing on the most informative transitions, outperforming existing subsampling techniques across various tasks and algorithms.

Offline reinforcement learning (RL) often involves repeatedly training policies from a fixed dataset of transitions. This process can be computationally expensive due to redundant data, while simple subsampling risks removing crucial rare transitions needed for complex, long-horizon tasks.CODS, or Critic-Guided Selector, addresses this by introducing an iterative approach to data selection. It alternates between fitting an algorithm-matched critic and acquiring transitions that exhibit high residuals, effectively identifying the most informative data points. This process results in a static, reusable subset of the original data.Unlike dynamic prioritized replay or one-shot selection methods, CODS refreshes its scores as the critic evolves, leading to a more refined and effective data subset. Benchmarking shows that CODS retains nearly all performance at a 10% data budget, significantly outperforming other baselines and demonstrating substantial compute advantages through data reuse.

Why it matters

For professionals developing and deploying offline RL systems, CODS offers a way to drastically reduce computational costs and training times while maintaining or improving policy performance, making RL more scalable and efficient.

How to implement this in your domain

  1. 1Investigate CODS for optimizing your offline reinforcement learning pipelines, especially with large datasets.
  2. 2Experiment with different data budget percentages to find the optimal balance between performance and computational savings.
  3. 3Integrate CODS into your RL training framework to create reusable, high-quality datasets.
  4. 4Analyze the impact of CODS on training stability and convergence for your specific RL tasks.

Original post by Ibne Farabi Shihab, Sanjeda Akter, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Md Najmus Swaqeeb, Anuj Sharma

"arXiv:2608.07719v1 Announce Type: new Abstract: Offline reinforcement learning repeatedly trains policies from a fixed transition pool, making redundant data costly across seeds and hyperparameters, while naive subsampling can remove rare transitions needed for long-horizon credi…"

View on X

Originally posted by Ibne Farabi Shihab, Sanjeda Akter, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Md Najmus Swaqeeb, Anuj Sharma on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses