We present MACT, a decentralized transformer-based world model designed for long-horizon multi-agent coordination while avoiding expensive planning. Each agent encodes its own discretized observation–action sequence using a shared transformer, while a Perceiver-style global module integrates cross-agent information under centralized training and decentralized execution. MACT achieves stable long-horizon coordination by combining the global latent representation with an action-conditioned contrastive objective that predicts future latent spaces multiple steps ahead. This allows agents to align their planned action windows with multi-step dynamics, producing more coherent rollouts and stronger team-level coordination. Experiments on the StarCraft II Multi-Agent Challenge (SMAC) demonstrate that MACT surpasses strong model-free baselines and previous world-model approaches on almost all tested maps, particularly in coordination-heavy scenarios.
Potential applications include robotics, autonomous multi-robot systems, and game AI agents. The ability to model long-horizon interactions among many agents can support industrial decision making, task allocation, and predictive control in complex real-world environments.
| 氏名 | コース | 研究室 | 役職/学年 |
|---|---|---|---|
| Victor Augusto Kich | システム科学コース | Learning Machines Group | 博士1回生 |