FedCars: An Efficient Scheduling Framework for In-Vehicle Federated Learning Integrated With CARLA-Cosimulation

Federated learning on vehicles has been widely considered in autonomous driving. The model uploading of federated learning is time-consuming, which motivates the transmission optimization that exploits the mobility of vehicles. In this paper, an optimization framework integrated with the high-fidelity traffic simulator CARLA, namely FedCars, is proposed for the scheduling of model uploading in a Vehicle-to-Infrastructure (V2I) network. Specifically, a group of data collection vehicles (DCVs) move along their planned routes with random velocities to collect data and train a model via federated learning, where the uplink transmission of multiple DCVs is jointly optimized. We first develop the FedCars simulator to obtain a high-fidelity trajectory dataset of DCVs in the considered traffic scenario, and train the transition probabilities of DCVs' locations. Hence, the uplink transmission time and power of all DCVs can be formulated as a finite-horizon Markov Decision Process (MDP), where the prediction of DCVs' future locations is exploited to minimize a weighted sum of the average model uploading time and energy consumption. A low-complexity FedCars optimizer with a non-trivial performance bound is then proposed, where the optimization consists of global transmission time and local power allocation. Since the former periodically allocates the transmission time for the remaining time slots, the power allocation of all DCVs can be decoupled into local finite-horizon MDPs which can be solved with low computational complexity. The simulation results show that the proposed solution framework can achieve both a significant performance gain over the baselines and a flexible trade-off between performance and complexity.