# Offline Q-Learning on Diverse Multi-Task Data Both Scales And Generalizes

**Authors:** Aviral Kumar, Rishabh Agarwal, Xinyang Geng, George Tucker, Sergey Levine  
**arXiv:** [2211.15144](https://arxiv.org/abs/2211.15144)  
**Date:** 2022-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.15144v1/

## Key points

- Aviral Kumar and colleagues explore the potential of offline reinforcement learning (RL) to create agents that generalize broadly
- High-capacity models trained on large, heterogeneous datasets can lead to such agents, similar to advances in vision and NLP
- ResNets, cross-entropy based distributional backups, and feature normalization are appropriate choices for scaling up model capacity in offline Q-learning algorithms
- A single policy trained on 40 games with near-human performance using up-to 80 million parameter networks scales favorably with capacity
- Offline Q-learning scales similarly with model capacity compared to return-conditioned supervised approaches but has better performance when the dataset is suboptimal
- Offline Q-learning with a diverse dataset is sufficient to learn powerful representations which facilitate rapid transfer to novel games and fast online learning on new variations of a training game
- The authors' contributions are significant: AK conceived and led the project; RA discussed the experiment design and project direction; XG helped with design choices for some experiments; GT advised the project and ran baseline DT experiments; SL advised the project and provided valuable suggestions. All authors contributed to writing and editing the paper.
- This work demonstrates promising results for scaling up offline RL methods while maintaining strong generalization capabilities which could have implications for developing more robust AI systems that can learn from diverse datasets and transfer their knowledge to new tasks.

## Summary

In their recent paper, "Offline Q-Learning on Diverse Multi-Task Data Both Scales And Generalizes," Aviral Kumar and colleagues explore the potential of offline reinforcement learning (RL) to create agents that generalize broadly. They argue that high-capacity models trained on large, heterogeneous datasets can lead to such agents, similar to advances in vision and NLP. To address the challenges of scaling up model capacity in offline RL methods, the authors re-examine previous design choices and find that with appropriate choices - ResNets, cross-entropy based distributional backups, and feature normalization - offline Q-learning algorithms exhibit strong performance that scales with model capacity. They test their approach using multi-task Atari as a testbed for scaling and generalization. The authors show that a single policy trained on 40 games with near-human performance using up-to 80 million parameter networks scales favorably with capacity. Furthermore, they demonstrate extrapolation beyond dataset performance even when trained entirely on a large but highly suboptimal dataset (51% human-level performance). Compared to return-conditioned supervised approaches, offline Q-learning scales similarly with model capacity and has better performance when the dataset is suboptimal. Additionally, they show that offline Q-learning with a diverse dataset is sufficient to learn powerful representations which facilitate rapid transfer to novel games and fast online learning on new variations of a training game. The authors' contributions are significant: AK conceived and led the project; RA discussed the experiment design and project direction; XG helped with design choices for some experiments; GT advised the project and ran baseline DT experiments; SL advised the project and provided valuable suggestions. All authors contributed to writing and editing the paper. Overall, this work demonstrates promising results for scaling up offline RL methods while maintaining strong generalization capabilities which could have implications for developing more robust AI systems that can learn from diverse datasets and transfer their knowledge to new tasks.