A Vision-Driven Evolutionary Game Framework for Distributed Task Allocation in Heterogeneous Multiagent Systems

This article proposes a vision-driven evolutionary game-theoretic framework for distributed task allocation in heterogeneous multiagent systems with minimum participation constraints. The framework integrates coalition formation with adaptive reinforcement learning, where agents dynamically adjust their aspiration levels through a heterogeneous vision function based on accumulated experience. We formulate the problem as a hedonic coalition formation game and design a fully distributed algorithm with three phases: initialization, vision-driven Q-learning, and postprocessing. Theoretical analysis proves that the game is an exact potential game, guaranteeing convergence to Nash-stable partitions under bounded rationality. We further establish an upper bound on the convergence rate using absorbing Markov chain theory and derive a finite-time regret bound for the learning process. Comprehensive simulations involving 100 agents and ten tasks demonstrate that our approach achieves 100% task completion within one iteration, reduces strategy oscillation by 85%, improves dynamic adaptation by over 70%, and maintains over 90% performance under sparse communication and sensing noise. Additional experiments on sensing noise, dynamic task insertion, scalability, and fairness confirm that vision-driven coalition formation maintains over 90% task completion under severe noise (2 m) and sparse communication, recovers from new tasks 60% faster than baselines, and achieves near-perfect fairness (Jain index 0.92) across heterogeneous agent types. The proposed method significantly outperforms the state-of-the-art baselines in convergence speed, stability, and robustness, offering a scalable solution for real-world cyber-physical systems.