Research map / Multi-agent systems and games / Games and equilibria
Learning to Optimize Differentiable Games
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learned optimizer L2PG instead of hand-designed SGA/ConOpt | two-player quadratic games · quadratic games transfer test (stable FPs) | convergence steps: 26.375 vs 35.963 (SGA) | Helped |
| Learned optimizer L2PG instead of SGA/ConOpt | four-player quadratic games · eps=2 transfer test | convergence steps: 25.00 vs 1000 (SGA, ConOpt) | Helped |
| Learned optimizer L2PG instead of SGA/ConOpt | GANs (2D mixture of Gaussians) · GAN training | convergence speed (qualitative) | Mixed |
| Learned optimizer L2PG on unstable-FP games | two-player quadratic games with unstable fixed points · unstable FPs test | convergence steps to unstable FP: 951.4 vs 968.1 (SGA) | No clear effect |
| Add Data-CL curriculum learning (ablation) | L2PG meta-training · quadratic games evaluation set | convergence steps: 54.313 vs 402.575 with no CL/SL | Helped |
| Add self-learning (SL) (ablation) | L2PG meta-training · quadratic games evaluation set | convergence steps: 91.113 vs 402.575 with no CL/SL | Helped |
| Add Training-CL (ablation) | L2PG meta-training · quadratic games evaluation set | convergence steps: 153.938 vs 402.575 with no CL/SL | Helped |
| Combine Data-CL, Training-CL, and SL (ablation) | L2PG meta-training · quadratic games transfer test | convergence steps: 26.375 vs 358.763 with no CL/SL | Helped |
| Use all three update components (cg, cA, cS) instead of only cg (ablation) | L2PG inputs and update coefficients · quadratic games transfer test | convergence steps: 26.375 vs 64.813 (only cg) | Helped |
| Use only the symmetric S g component (no gradient/antisymmetric) (ablation) | L2PG components · quadratic games transfer test | convergence steps: 372.263 vs 26.375 (all components) | Hurt |
| Multiplicative stability-aware loss vs additive regularization vs no regularization (ablation) | L2PG meta-training · toy game with unstable FP (Figure 4) | convergence to unstable FP | Helped |
Closest papers by meaning
- The Mechanics of n-Player Differentiable Games (ICML 2018 · similarity 0.89)
- Extra-gradient with player sampling for faster convergence in n-player games (ICML 2020 · similarity 0.87)
- Generative Adversarial Equilibrium Solvers (ICLR 2024 · similarity 0.86)
- Learning Equilibria in Adversarial Team Markov Games: A Nonconvex-Hidden-Concave Min-Max Optimization Problem (NeurIPS 2024 · similarity 0.86)
- Constrained Exploitability Descent: An Offline Reinforcement Learning Method for Finding Mixed-Strategy Nash Equilibrium (ICML 2025 · similarity 0.86)
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.86)
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.