Semantic Communications for Partially Observable Multi-Agent Reinforcement Learning-Based Unmanned Aerial Vehicles Monitoring System

In multi-agent reinforcement learning (MARL) environments with partial observability, agents lack access to a global state, limiting their decision-making capabilities. To address this, agents can share local observations via inter-agent communications; however, in practice, transmission failure can happen in unreliable channel conditions, e.g., packet loss, thereby degrading MARL performances. This paper proposes a semantic communication framework that enhances MARL in partially observable environments, specifically for the UAV-based disaster monitoring system facing with unreliable communications.