Constrained Multi-Agent Reinforcement Learning for Task Offloading in Wireless Edge Networks

In edge computing systems, autonomous agents must make fast local decisions while competing for shared resources. Existing MARL methods often assume centralized critics or frequent communication, which fail under limited observability and communication constraints. We propose a decentralized framework in which each agent solves a constrained Markov decision process (CMDP), coordinating implicitly through a shared constraint vector that limits the frequency of offloading. These constraints are updated infrequently and act as a lightweight coordination mechanism, enabling agents to align with global resource usage objectives with little direct communication. Using safe reinforcement learning, agents learn policies that meet both local and global goals. We establish theoretical guarantees under mild assumptions and validate our approach experimentally, showing improved performance over centralized and independent baselines, especially in large-scale settings.