Federated learning for vehicle trajectory prediction

Time Series data has become present everywhere, thanks to affordable edge devices and sensors. Much of this data is valuable for decision making. To use this data for the forecasting task, the conventional centralized approach has shown deficiencies regarding extensive data communication and data privacy issues. Federated learning allows learning from decentralized data without the need to store it centrally. The data remains where it was generated, which guarantees privacy and reduces communication costs [1]. FL already naturally selects only a few nodes randomly at each round. They have non-iid data and also varying in amount. After some iteration of training, the central aggregator will generate a global model. That is, heavily learning depends on a coordinator, which causes scalability issues with large numbers of nodes, and besides, there is a single point of failure, which is not suitable for some applications. An example of this is predicting the online trajectory of moving nodes to manage traffic, or proactive resource allocation in vehicular networks. In this work, to tackle these problems, we use personalized distributed Federated Learning which is online, peer-to-peer and provides asynchronous communications. Each node in this network is a client for other existing nodes and use its local dataset to improve their models. At the same time, it is like a coordinating server that merges received models and personalized the model for itself. There can be as many models as many as the number of clients. We present three practical algorithms called DFed Avg, DFed Pow and DFed Best for the serverless federated learning of deep networks based on iterative model averaging, and an empirical evaluation which considers time series datasets and an LSTM model. DFed Avg merges models based on the technique used in Federated Averaging, while DFed Best, and DFed Pow at every iteration rounds use different methods to merge models. Our goal is to evaluate our optimization algorithms, not to achieve the best possible accuracy on these tasks. The experiments demonstrate that these approaches are rather robust and can have numerous clients with the dynamic, unbalanced and non-IID data distributions.