Collaborative Decision-Making in AI Multi-Agent Systems

Abstract

This study investigates collaborative decision-making in multi-agent systems (MAS) operating in dynamic environments, emphasizing cooperative behavior and reinforcement feedback. Unlike centralized planning models, the proposed approach allows agents to communicate adaptively, aligning individual actions with shared objectives without a global controller. Using a custom simulation in the PettingZoo library and Multi-Agent Reinforcement Learning (MARL), agents were trained to perform resource gathering and conflict avoidance in a grid-based setting. Evaluation across different agent densities and resource configurations revealed that agents with structured communication and shared goals outperform independent agents in task completion, coordination, and collision avoidance. Empirical results, including tabular and graphical data, demonstrate the advantages of collaborative decision-making. The study also addresses challenges such as scalability, convergence delays, and communication overhead, laying the foundation for real-world applications in autonomous transport, disaster response, and smart infrastructure systems.