MAIT: Multi-agent Local Observation Interaction to Improve the Decision-Making Ability

In multi-agent systems, partial observability significantly limits the acquisition of optimal joint actions. The main reason is the limitations in each agent's perception of the overall environment, enabling them to perform only optimal actions based on localized information. However, the combination of these actions may result in suboptimal joint actions, causing multi-agent systems to fall into local optimal solutions. Addressing this challenge, this paper proposes an innovative universal multi-agent deep reinforcement learning framework called MAIT (Multi-Agent Interact Transformer). The core of the MAIT framework is to enhance the understanding of the global environment by interactively sharing local observation information among agents. The framework utilizes the Attention mechanism, tightly integrating agents’ observations with their actions and assigning weights based on importance. This approach not only strengthens the agents’ perception of the environment but also enhances the efficacy of their actions. Furthermore, to address the issue of vast observation and action spaces in multi-agent reinforcement learning, we introduced an Additive Attention mechanism for lightweight improvements, enhancing algorithm efficiency. Experiments conducted in the SMAC (StarCraft Multi-Agent Challenge) environment show that MAIT can enhance agents’ ability to recognize the global environment through observational interactions, achieving more effective collaboration and significantly improving algorithm performance, especially in challenging tasks.