LLT: Learning Latent Tactics in Uncertainty Cooperative Multi-Agent Reinforcement Learning

State-of-the-art multi-agent reinforcement learning (MARL) algorithms typically follow the Centralized Training with Decentralized Execution (CTDE) paradigm, where agents independently decide based on their observations during execution. However, human players in strategic games like StarCraft II highlight the critical role of timely information sharing and coordinated tactics. Inspired by this, we propose Learning Latent Tactics (LLT) in cooperative MARL, a novel collaboration framework. During training, we jointly optimize the agents’ policy networks and a tactical decision model. In execution, agents select tactics based on comprehensive information, including that of their teammates, and execute actions grounded in both the selected tactic and their policy networks. Notably, our approach is resilient to scenarios of uncertainty cooperative multi-agent reinforcement learning, where some or even all of the other agents’ information may be temporarily unavailable, making it suitable for real-world open collaboration settings. Experimental results on the StarCraft II game show that our approach outperforms the baseline, and ablation studies validate the effectiveness of our LLT collaboration framework.