Online Model Retraining and Instance Allocation in Edge Computing Networks

The adoption of deep learning models in V2X scenarios has boosted computing demands in edge computing, while concept drift requires frequent model retraining, further increasing computing resource consumption. However, few works study computing instance allocation considering dynamic model retraining, especially under varying workloads. In this paper, we study the joint online model retraining and instance allocation problem in edge computing networks considering model performance degradation due to concept drift. Solving the online problem is challenging since it is a quadratic binary programming problem and its instance switching cost is time-coupling. We propose an efficient online algorithm, where we first linearize the quadratic term, then regularize the time-linearize the problem and then regularize the timecoupling switching cost to decouple the problem, and finally round the fractional solution by a randomization method. We prove that our proposed algorithm achieves a bounded optimality gap. Simulations demonstrate that our algorithm can achieve a balance between instance costs and model performance.