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

arxiv:2305.19452

Original source · Paper JSON

Reinforcement learning · 2023-05-30

Cites (11) · Cited by (3)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Bigger, Better, Faster: Human-level Atari with human-level efficiency
• Abstract
• Keywords:
• 1 Introduction
• 2 Background
• 3 Related Work
• 4 Method
• Base agent.
• Harder resets.
• Larger network.
• Receding update horizon.
• Increasing discount factor.
• Weight decay.
• Removing noisy nets.
• 5 Analysis
• The importance of self-supervision.
• Sample efficiency via more gradient steps.
• The surprising importance of target networks
• Reset Strength
• Scale is not enough on its own.
• Computational efficiency.
• 6 Revisiting the Atari 100k benchmark
• New Frontiers
• Data Scaling
• 7 Discussion and Future Work
• Appendix A Additional Results
• Instructions for reporting errors