ADAPT-ROMA: Adaptive Risk-Aware Multi-Objective Offloading for Robust Mobile Edge Computing

Mobile Edge Computing (MEC) offers significant advantages in enhancing user experience and reducing energy consumption by offloading tasks to proximate edge servers. However, MEC systems are inherently complex, facing highly dynamic environments with fluctuating channel conditions and server loads, coupled with the necessity to optimize multiple conflicting objectives such as latency and energy consumption. Furthermore, existing offloading strategies often lack robustness against inherent uncertainties, leading to performance degradation in unpredictable scenarios. To address these critical challenges, this paper introduces ADAPT-ROMA, a novel Adaptive Dynamic Risk-aware Multi-objective Offloading Algorithm. We formulate the MEC offloading problem as a Contextual Risk-aware Multi-objective Markov Decision Process, integrating Conditional Value at Risk (CVaR) to explicitly quantify and mitigate worst-case performance losses. ADAPT-ROMA employs an adaptive weight generation network to dynamically balance latency, energy, and risk objectives based on real-time system states. Its core is built upon a robust distributed deep reinforcement learning framework, enhanced by adversarial training to bolster resilience against unmodeled uncertainties, and a hierarchical attention network for effective contextual encoding. Extensive simulations demonstrate that ADAPT-ROMA achieves a superior balance between average task latency and total energy consumption, alongside an outstanding task completion rate. Crucially, it yields a significantly low risk-aware score, outperforming baseline methods by ensuring high performance and stability even under dynamic and uncertain MEC conditions. Ablation studies, adaptability analysis, robustness evaluation, and scalability tests further confirm the individual contributions of ADAPT-ROMA’s components and its practical applicability.