Адаптивна мультиагентна система з динамічним перемиканням стратегій у змагально-кооперативному ігровому середовищі

Abstract

This paper presents an adaptive multi-agent reinforcement learning (MARL) system for a mixed cooperative-competitive game environment. The proposed model introduces the Adaptive Mode Switching (AMS) mechanism, which allows agents to dynamically shift between cooperative and competitive behavior depending on the current environment state. A custom grid-based game environment was developed using the Gymnasium API, featuring partial observability and resource control zones. The AMS meta-signal is integrated into the MADDPG reward function, enabling agents to independently learn when cooperation or competition yields greater benefit without explicit switching rules.