A Feedback Control Framework Approach to Federated Learning With Adaptive Optimizer

Federated learning (FL) reaps high accuracy by protecting data ownership, governance or privacy. Compared to conventional Federated Averaging (FedAvg), Federated Learning with Adaptive Optimizer (FedAdam) achieves better convergence performance by dynamically adjusting both optimization direction and learning rate. However, FedAdam suffers from impulse overshooting which induces poor convergence performance in terms of accuracy and/or convergence speed. To improve the convergence performance of FedAdam, we first borrow the idea of feedback control to re-interpret FedAdam as an optimizer consisting only of integral term (I). Next, we propose an adaptive federated optimization algorithm by incorporating proportional term P and differential term D to form a centralized PID controller at the parameter server. The proposed FedAdamPID generalizes FedAdam while the paper consider the problem of hyperparameters P, I and D optimization. In particular, we propose two methods to find the values of P, I, and D terms, including grid searching over three dimensions (Grid), and emloying reinforcement learning, namely deep deterministic policy gradient (DDPG) The two schemes are called Grid-FedAdam-PID and DDPG-FedAdam-PID, respectively. A series of experimental tests on typical models and datasets for Internet of Things (IoT) settings verified that FedAdam-PID achieves significantly better convergence performance when compared with FedAdam. DDPG-FedAdam-PID provides the best performance, compared with grid search, while also being flexible enough for different environments.