Computation Offloading for Distributed Learning in Vehicular Networks: A Service Scheduling and Resource Allocation Method

Distributed machine learning has attracted significant attention for the future vehicular networks. However, the inadequate computation capacity of multi-vehicles and the dynamic wireless communication environment degrade the training effectiveness of distributed models. Against these backdrops, a novel end-edge-cloud collaborative computation offloading for distributed learning paradigm is proposed in this paper. In order to minimize the weighted sum cost of latency and learning loss, we jointly optimize the offloading decisions for access selection and the proportion of data offloading, as well as the resource allocation strategy for bandwidth and transmit power. Wherein, the straggler effect is addressed by formulating a dynamic short timescale problem based on service fairness and a combinatorial iterative approach is suggested to resolve this min-max problem. Then, over the long timescale, the modified multi-agent reinforcement learning (MARL) algorithm based on multiple base stations and users is developed to overcome the dilemma of the missing explicit relationship between optimization variables and learning loss function. Moreover, a fractional programming (FP) based MARL algorithm is designed to solve this NP-hard issue. Simulation results verify the superiority of the proposed FP-MARL algorithm in comparison to the benchmark schemes.