Controlling Heterogeneous Multi-Agent Systems Under Uncertainty Using Fuzzy Inference and Evolutionary Search

Real-time coordination of heterogeneous multi-agent systems in dynamic and partially observable environments poses significant challenges. To address this, we propose a framework that integrates fuzzy inference systems with real-valued genetic algorithms to optimize decision-making under strict time constraints and sensory uncertainty. We evaluate the proposed method in the RoboCup Soccer Simulation 2D League, where 22 autonomous agents coordinate through a fuzzy-evaluated action sequence search. Spatial heuristics are encoded as fuzzy rules, and optimization based on genetic algorithms refines evaluation function parameters according to performance metrics such as number of shots, goal area entries, and scoring rates. The resulting control strategy remains interpretable; spatial heat maps reveal emergent behaviors such as coordinated positioning and ridgeline passing patterns near the penalty area. The experiments against established RoboCup teams, serving as benchmarks, demonstrate the competitive performance of our trained agents while enabling analyses of evolving decision structures and agent behaviors. Our method provides a transparent and adaptable framework for controlling heterogeneous agents in uncertain real-time environments, with broad applicability to robotics, autonomous systems, and distributed control systems.