Research map / Multi-agent systems and games / Games and equilibria
Scaling Laws for a Multi-Agent Reinforcement Learning Model
Oren Neumann, Claudius Gros
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase neural network parameter count | AlphaZero · Connect Four and Pentago | playing strength: α_N: Connect Four $0.88$; Pentago $0.87$ | Improved |
| Increase training compute while using optimally sized agents | AlphaZero · Connect Four and Pentago | playing strength: α_C: $0.55$ for both games | Improved |
| Scale model size with the available compute budget | AlphaZero · Connect Four and Pentago | optimal model size: α_C^opt: Connect Four $0.62$; Pentago $0.63$ | Improved |
| Use a larger neural network at the same amount of training data | AlphaZero · Connect Four and Pentago | Elo score | Improved |
| Increase model size on Oware | AlphaZero · Oware | Elo score: α_N $0.75$ | Mixed |
| Increase model size on Oware | AlphaZero · Oware | Elo score | Hurt |
| Use test-set value or policy loss as the performance indicator instead of Elo | AlphaZero · Connect Four solver-generated test set | performance prediction: 0.336 | No clear effect |
| Train beyond the compute-optimal point until convergence | AlphaZero · Connect Four | Elo score per fixed compute budget | Hurt |
Closest papers by meaning
- AlphaZero Neural Scaling and Zipf's Law: a Tale of Board Games and Power Laws (NeurIPS 2025 · similarity 0.93)
- Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- A Dynamical Model of Neural Scaling Laws (ICML 2024 · similarity 0.88)
- Mixtures of Experts Unlock Parameter Scaling for Deep RL (ICML 2024 · similarity 0.88)
- Bigger, Regularized, Optimistic: scaling for compute and sample efficient continuous control (NeurIPS 2024 · similarity 0.87)
- Network Sparsity Unlocks the Scaling Potential of Deep Reinforcement Learning (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
