Bayesian Deep Neural Network-empowered Thompson Sampling for Context-aware Task Offloading in Dynamic Fog Computing

Efficient task offloading in dynamic fog computing environments requires adaptive decision-making under uncertainty. This paper proposes a Bayesian Deep Neural Network (BDNN)-empowered Thompson Sampling (TS) framework for context-aware task offloading, enabling intelligent resource allocation while balancing exploration and exploitation. The BDNN models the stochastic reward function by learning a posterior distribution over network weights, capturing the uncertainty in offloading decisions. At each time step, the task node samples a set of weights from the learned posterior to estimate the expected reward of each helper node, facilitating adaptive decision-making in dynamic network conditions. Experimental results demonstrate that our approach outperforms conventional heuristics and deep learning-based methods, achieving lower latency, improved resource utilization, and better offloading efficiency in fog computing environments.