Research map / Multi-agent systems and games / Games and equilibria
Adversarial Policy Learning in Two-player Competitive 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 |
|---|---|---|---|
| Use value-difference objective (maximize adversary value minus victim value, monotonic surrogate) instead of PPO on adversary reward alone | adversarial agent training · StarCraft II | adversary winning rate: 98% vs 31% | Helped |
| Use value-difference objective instead of PPO on adversary reward alone | adversarial agent training · MuJoCo games (You-Shall-Not-Pass, Kick-And-Defend, Sumo-Humans) | adversary winning rate: higher winning rate in all three games | Helped |
| Use value-difference objective instead of PPO on adversary reward alone | adversarial agent training · Sumo-Ants | adversary winning rate | No clear effect |
| Use value-difference objective instead of PPO on adversary reward alone | adversarial agent training · Sumo-Ants | non-loss rate: about 88% non-loss for proposed | Helped |
| Use value-difference objective instead of PPO on adversary reward alone | adversarial agent training · StarCraft II | victim win plus tie rate after blinding (adversary reliance on observation): 24% vs 98% for baseline after blinding | Mixed |
| Retrain victim against the proposed adversary (adversarial retraining) | victim agent · StarCraft II | victim win plus tie rate vs baseline adversary: 89% vs 4% for baseline-retrained victim against our adversary | Helped |
| Retrain victim with the proposed adversary instead of baseline adversary | victim agent · Kick And Defend | victim win plus tie rate vs our adversary: 68.0 vs 35.0 | Helped |
| Retrain victim against adversary; check regular-agent performance | victim agent · regular agents (Kick And Defend, You Shall Not Pass, Sumo Humans, Sumo Ants, StarCraft II) | win rate vs regular agents: negligible influence | No clear effect |
| Drop monotonic property (no-monotonic variant) of the value-difference objective (ablation) | adversarial agent training · not reported quantitatively | adversary winning rate | Hurt |
| Use only victim-reward term as objective instead of value difference (ablation) | adversarial agent training · not reported quantitatively | adversary winning rate | Hurt |
Closest papers by meaning
- Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL (ICLR 2024 · similarity 0.90)
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.90)
- Adversaries With Incentives: A Strategic Alternative to Adversarial Robustness (ICLR 2025 · similarity 0.88)
- Off-Policy Actor-Critic for Adversarial Observation Robustness: Virtual Alternative Training via Symmetric Policy Evaluation (ICML 2025 · similarity 0.87)
- Representation-Based Robustness in Goal-Conditioned Reinforcement Learning (AAAI 2024 · similarity 0.87)
- RAT: Adversarial Attacks on Deep Reinforcement Agents for Targeted Behaviors (AAAI 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.