# Bigger, Better, Faster: Human-level Atari with human-level efficiency

**Authors:** Max Schwarzer, Johan Obando-Ceron, Aaron Courville, Marc Bellemare, Rishabh Agarwal, Pablo Samuel Castro  
**arXiv:** [2305.19452](https://arxiv.org/abs/2305.19452)  
**Date:** 2023-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.19452v3/

## Key points

- Authors introduced a value-based RL agent named BBF that surpasses super-human performance in the Atari 100K benchmark
- Success of BBF attributed to scaling of neural networks for value estimation and strategic design choices for efficient sample utilization
- Extensive analyses of design decisions offer valuable insights for future research in reinforcement learning
- Discussion on redefining benchmarks for sample-efficient RL research on the ALE platform
- Code and data openly accessible at https://github.com/google-research/google-research/tree/master/bigger_better_faster

## Summary

In their paper titled "Bigger, Better, Faster: Human-level Atari with human-level efficiency," authors Max Schwarzer, Johan Obando-Ceron, Aaron Courville, Marc Bellemare, Rishabh Agarwal, and Pablo Samuel Castro introduce a value-based RL agent named BBF that surpasses super-human performance in the Atari 100K benchmark. The success of BBF is attributed to the scaling of neural networks for value estimation and several other strategic design choices that facilitate efficient sample utilization. Through extensive analyses of these design decisions, the authors offer valuable insights for future research in reinforcement learning. Additionally, they conclude their study with a discussion on redefining benchmarks for sample-efficient RL research on the ALE platform. The authors have made their code and data openly accessible at https://github.com/google-research/google-research/tree/master/bigger_better_faster. This work was presented at ICML 2023 and is a revised version of the original publication.