RS-PLO: Risk-Sensitive Predictive Lyapunov Optimization for Dynamic Task Offloading in Multi-Edge Computing Environments

Abstract

Mobile Edge Computing (MEC) has emerged as a transformative paradigm that brings computation closer to the network edge, enabling resource-constrained IoT devices to offload computation-intensive tasks to nearby edge servers and thereby significantly reducing end-to-end latency and energy consumption. Existing MEC frameworks predominantly rely on static optimization parameters that cannot respond to the non-stationary channel dynamics caused by device mobility, interference, or sudden obstructions, leading to queue instability and degraded system performance. This paper presents RS-PLO (Risk-Sensitive Predictive Lyapunov Optimization), a dynamic task offloading framework that introduces a volatility queue to monitor channel prediction errors in real time and an adaptive control parameter that decays exponentially with accumulated channel uncertainty, automatically balancing energy optimization against queue stability. The framework incorporates multi-edge server selection with fairness-aware allocation, drift-corrected channel prediction, and a risk-gated admission controller that defers high-complexity tasks during volatile channel periods. Experimental evaluation across five representative scenarios demonstrates a 4.0% improvement in queue stability, a 5.4% reduction in task latency, up to 18.1% improvement in channel prediction accuracy, and more selective offloading behavior compared to static Lyapunov baselines, all without manual parameter tuning. Statistical significance is confirmed across ten independent trials (p < 0.05, paired t-test)..