Moving Edge for On-Demand Edge Computing: An Uncertainty-aware Approach
Fangtong Zhou, Ruozhou Yu
TL;DR
The paper tackles the problem of fluctuating edge demands by introducing Moving Edge resources (MUs) and a learning-based, uncertainty-aware scheduling framework called URANUS. It combines Bayesian neural network-based demand prediction with distributionally robust post-processing (WC-CVaR) and finite-horizon planning to optimize MU deployment under data and model uncertainty while respecting SLA penalties. Key contributions include a BVNN-based time-series predictor with uncertainty quantification, a distributionally robust CVaR-based penalty estimation reformulated as a tractable SDP, and a corresponding single- and multi-agent finite-horizon planning algorithm. Experiments on real Milan data show URANUS substantially improves robustness and profit versus end-to-end reinforcement learning, uncertainty-agnostic methods, and heuristic baselines, demonstrating practical potential for on-demand edge computing using moving modular data centers.
Abstract
We study an edge demand response problem where, based on historical edge workload demands, an edge provider needs to dispatch moving computing units, e.g. truck-carried modular data centers, in response to emerging hotspots within service area. The goal of edge provider is to maximize the expected revenue brought by serving congested users with satisfactory performance, while minimizing the costs of moving units and the potential service-level agreement violation penalty for interrupted services. The challenge is to make robust predictions for future demands, as well as optimized moving unit dispatching decisions. We propose a learning-based, uncertain-aware moving unit scheduling framework, URANUS, to address this problem. Our framework novelly combines Bayesian deep learning and distributionally robust approximation to make predictions that are robust to data, model and distributional uncertainties in deep learning-based prediction models. Based on the robust prediction outputs, we further propose an efficient planning algorithm to optimize moving unit scheduling in an online manner. Simulation experiments show that URANUS can significantly improve robustness in decision making, and achieve superior performance compared to state-of-the-art reinforcement learning, uncertainty-agnostic learning-based methods, and other baselines.