Research map / Multi-agent systems and games / Games and equilibria
Feint Behaviors and Strategies: Formalization, Implementation and Evaluation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formalizes feint behaviors at action and strategy level (palindrome-directed templates, dual-behavior model, temporal/spatial/collective rewards) and implements them in MARL training for boxing games.
Problem
Existing literature lacks a comprehensive, concrete formalization of feint behaviors and their strategic implications in multi-player competitive games.
Approach
Generates feint behaviors from palindrome-like (semi-symmetric) subsequences of attack behaviors, composes them with follow-up high-reward actions via a Dual-Behavior Model, and defines temporal (short-long-term), spatial (occupancy-measure divergence) and collective reward terms that are added to MARL training as a separate feint policy.
Key results
- Feint-equipped agents gain a significant reward advantage over baseline opponents in the two-player boxing game across all four MARL models (MADDPG, MASAC, MATD3, MAD3PG).
- In the six-player scenario, the Feint-equipped player earns significantly more reward than both opponents and its teammates across all four MARL models.
- Feint-enabled agents reach around 4.9e-2 Exploitability versus about 9.7e-2 for regular agents, and the paper claims a 1.98X exploitation gain; overhead is reported as under 5% per episode.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Feint behaviors to one player | MADDPG, MASAC, MATD3, and MAD3PG · Boxing game, 1 vs 1 and 3 vs 3 | per-episode game reward | Helped |
| Add Feint actions | AlphaStar policies · AlphaStar | exploitability: around $9.7 \times 10^{-2}$ -> around $4.9 \times 10^{-2}$; 1.98X | Helped |
| Add Feint actions | AlphaStar policies · AlphaStar | negative PE: around $1.2 \times 10^{-2}$ -> around $5.3 \times 10^{-2}$ | Helped |
| Add Feint to MARL training | Feint implementation · Boxing game | per-game-episode time consumption: less than 5\% | Hurt |
| Add Feint actions | AlphaStar policies · AlphaStar | variation in exploitability and negative PE | Hurt |
Surprising findings
- MASAC advantage for the Feint player is unstable and the authors attribute this to MASAC's own instability.
Limitations
- Evaluation uses a custom handcrafted boxing simulator rather than a standard benchmark, with rendering omitted.
- Feint is only evaluated in boxing and a modified AlphaStar setting; generalization to other sports is claimed but not tested.
- Several hyperparameters (lambda, mu weights) are hand-set and only described as adjusted during training.
Experimental setup
Datasets: Mixamo boxing behaviors (22 behaviors, over 250 actions)
Benchmarks: custom 1v1 boxing game custom 3v3 (six-player) boxing game AlphaStar (diversity analysis)
Metrics: per-episode game reward Exploitability Population Efficacy running time overhead
Baselines: MADDPG MASAC MATD3 MAD3PG
Builds on: MADDPG MASAC MATD3 MAD3PG Multi-Agent Particle Environment
Closest papers by meaning
- A Causal World Model Underlying Next Token Prediction: Exploring GPT in a Controlled Environment (ICML 2025 · similarity 0.82)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.82)
- The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2026 · similarity 0.82)
- Opponent Modeling based on Subgoal Inference (NeurIPS 2024 · similarity 0.82)
- Implicit Search via Discrete Diffusion: A Study on Chess (ICLR 2025 · similarity 0.82)
- Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.82)
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.