Constrained hybrid soft actor–critic (CH-SAC) with a mobility-and load-aware fusion predictor

Abstract

Fog computing provides a distributed execution layer for latency-sensitive and mobility-driven applications, but its effectiveness is constrained by strong spatiotemporal demand variability, bursty arrivals, and congestion feedback. In realistic urban environments, reactive controllers and mobility-only predictors often respond only after queues have formed, causing sustained congestion, inflated tail latency, packet drops, and Quality-of-Service (QoS) degradation. These limitations are especially severe under compute bottlenecks, constrained buffering, and rapidly shifting demand hotspots. This thesis proposes CH-SAC with a Mobility- and Load-Aware Fusion Predictor for proactive fog-resource provisioning. The framework combines a Flow-GRU mobility head with a Holt-style offered-load demand head that operates on accepted plus dropped requests, and fuses them through adaptive congestion-aware weighting with explicit floor and cap safety bounds. The resulting predictive signals are integrated into a Constrained Hybrid Soft Actor–Critic (CH-SAC) controller supported by a guard mechanism that monitors queue pressure, utilisation, latency spikes, and drop ratio. The framework is evaluated using real mobility traces from the Beijing T-Drive dataset in both a trace-driven YAFS environment and a Mininet-based fog emulation testbed under CPU-stress, queue-stress, and flash-crowd / hotspot-shift conditions. The proposed framework forms a closed online control loop in which mobility-driven workload is forecast at a short horizon, hotspot nodes are prioritised proactively, and continuous resource-scaling actions are issued before congestion becomes dominant. In the YAFS behavioural analysis, the Fusion Predictor achieves strong short-horizon accuracy (MAE = 3.45 taxis, Pearson correlation r ≈ 0.95). Under a proactive-action definition based on scaling before overload onset, the controller attains a proactive-action share of 78.2%, a median lead time of 17.0 s, and an overload prediction F1- score of 0.890. In the Mininet stress evaluation, CH-SAC with Fusion Predictor achieves 79.3 ms mean latency and 241,454 completed requests under queue stress, and 41.65 ms mean latency with 162,285 completed requests under flash-crowd / hotspot-shift stress. Across the two evaluation layers, the framework improves proactive provisioning, latency, congestion control, service reliability, and recovery behaviour relative to threshold-based, metaheuristic, baseline RL, and non-predictive SAC baselines. The results also clarify the complementary roles of the framework’s two main enhancements: the Fusion Predictor is especially important when early anticipation is required, particularly under CPU- and queue-stress conditions, while the guard is essential for robustness and survivability under abrupt flash-crowd stress. Overall, the thesis shows that integrating fused short-horizon forecasting with constraint-aware continuous-control reinforcement learning transforms fog-resource provisioning from reactive overload response into proactive, resilient, and QoS-oriented control in mobility-driven urban environments.

Date	20 Jun 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Michel Kadoch (Supervisor)