LyMARL: Lyapunov-Guided MARL for Energy-Constrained User Association

초록

We investigate the joint optimization of user association (UA) and base station (BS) activation in multi-cell wireless networks under finite-horizon energy constraints. This problem is challenging due to the inherent tension between maximizing user throughput and ensuring energy-efficient BS operation. Lyapunov-based control provides a principled and practical approach for distributed stochastic network optimization with strong long-term guarantees; however, energy constraints are enforced only in an asymptotic sense, which may lead to noticeable violations over finite horizons. To address this limitation, we propose LyMARL, a Lyapunov-guided multi-agent reinforcement learning framework that jointly optimizes UA and BS activation and scheduling by embedding Lyapunov virtual queues into agent observations and reward design. Unlike conventional Lyapunov-based approaches with passive BS control, LyMARL introduces proactive BS agents for energy-aware decision making while preserving fully decentralized execution. Simulation results demonstrate that LyMARL achieves substantial improvements in throughput, fairness, and energy efficiency.

키워드