JumpStart Your Policy Learning with Lessons from 160,000 Training Runs
A study of 160,000 offline RL and imitation policies finds no dominant algorithm and releases JumpStart.
Researchers report a large-scale study of offline reinforcement and imitation learning, training more than 160,000 policies across 114 datasets. No algorithm dominates: the strongest methods are often close in aggregate, but leaders change by environment, and hyperparameter tuning frequently reshuffles rankings. Benchmark composition can produce conflicting conclusions. They describe a simple strategy for strong default hyperparameters and a dataset-conditioned recommender, and release JumpStart with trained policies, scores, hyperparameters, baselines, code, and a results website.
- More than 160,000 policies were trained across 114 datasets.
- No single algorithm leads across all environments.
- Hyperparameter tuning and benchmark mix can reverse rankings.
- JumpStart releases policies, scores, code, baselines, and a recommender.
- Products
- JumpStart
From the original article
1 min read at huggingface.co
Reliable progress in offline policy learning depends on careful reporting, well-tuned baselines, and evaluation across diverse conditions. Prior work has shown that results can be sensitive to reporting choices, hyperparameter tuning, and dataset properties, but these sources of variability have not been systematically investigated together at the scale needed to understand how they shape conclusions.
By Nabil Omi, Eric Bae, Chung Yik Edward Yeung · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.