FedSysID: A Federated Approach to Sample-Efficient System Identification

We study the problem of learning a linear system model from the obse rvations of M clients. The catch: Each agent is observing data from a different dynamical system. This work addresses the question of how multiple systems collaboratively learn dynamical models in the presence of heterogeneity. We pose this problem as a federa ted learning problem and characterize the tension between achievable performance an d system heterogeneity. Furthermore, we provide a sample complexity result that obtains a c onstant factor improvement over the single agent setting. Finally, we describe a meta federated learning algorithm, FedSysID, that leverages existing federated algorithms at the client level.