A predictive multi-agent learning framework for energy-constrained ultra-reliable edge systems under stochastic environments

Ultra-reliable low-latency systems operating at the network edge are increasingly constrained by stochastic environmental factors, including energy variability, dynamic workloads, and time-varying communication conditions. Existing learning-based resource management approaches predominantly rely on reactive decision-making and assume stable energy availability, limiting their applicability in energy-harvesting edge systems where uncertainty is intrinsic and future conditions critically influence system reliability. This study proposes a predictive multi-agent reinforcement learning framework for energy-constrained edge systems, designed to enable proactive and resilient decision-making under stochastic dynamics. The framework integrates a Transformer-enhanced graph attention predictor with a decentralized multi-agent learning architecture based on centralized training and decentralized execution. By jointly modeling energy harvesting processes, workload dynamics, and communication variability, the proposed approach allows agents to anticipate future system states and adapt resource allocation policies accordingly. A key contribution is the introduction of an energy-deficit-aware reward shaping mechanism, which incorporates both current and predicted energy conditions to guide agents toward proactive coordination and stability preservation. This mechanism enables emergent cooperative behavior without explicit inter-agent communication, improving robustness in non-stationary environments. Extensive simulations using trace-driven network environments demonstrate that the proposed framework significantly improves system reliability, energy efficiency, and latency performance compared to state-of-the-art reactive and optimization-based baselines. In particular, the approach reduces energy depletion events, stabilizes queue dynamics, and maintains ultra-reliable performance under severe environmental fluctuations.