FedForce: Network-adaptive Federated Learning for Reinforced Mobility Prediction

Federated Learning (FL) has become popular in the field of mobility and trajectory prediction due to its privacy-preserving and scalability capabilities. Deploying FL over resource-constrained devices and varying network conditions is challenging for achieving a good tradeoff among prediction performance, computational load, and communication volume. On the other hand, the design of FL’s distributed neural architectures is complex, time-consuming, and dependent on experts’ prior knowledge. To tackle the above limitations, we propose the network-adaptive FEDerated learning for reinFORCEd mobility prediction (FedForce) system. FedForce employs reinforcement learning to design a transformer neural network whose architecture jointly optimizes the prediction accuracy, training time, and transmission time based on the mobility dataset’s unique features, the client’s computing capacity, and the available network throughput. FedForce outperforms several state-of-theart trajectory predictors and achieves an average displacement error of 0.20m on the ETH+UCY dataset and an accuracy of 76% on the Orange dataset (-0.02m and 10% higher than the bestperforming baseline, respectively), while cutting the FL training and transmission time by half. FedForce can save up to 80% of computational resources and 96% of communication overheads with a negligible accuracy decrease.