1: \begin{abstract}

2: %    Federated averaging (FedAvg) is a popular federated learning (FL) technique that updates the global model by averaging local models and then transmits the updated global model to devices for their local model update. One main limitation of FedAvg is that the average-based global model is not necessarily better than local models so that FedAvg might degrade in realistic scenarios, especially when the data is non-identically distributed across devices and the number of data samples varies significantly from device to device. In this paper, we propose a technique that improves the quality (accuracy and convergence speed) of FL greatly. The key idea of the proposed technique, henceforth referred to as \textit{simulated annealing-based FL} (SAFL), is to allow a device to choose its local model instead of the global model. Specifically, by exploiting the simulated annealing strategy, SAFL makes each device choose its local model with high probability in early iterations when the global model is immature. Whereas, when the global model becomes mature after some iterations, the global model is chosen with high probability. From extensive numerical experiments based on various benchmark datasets, we demonstrate that SAFL outperforms the conventional FedAvg technique in terms of the convergence speed and the classification accuracy.

3: %%For example, SAFL converges two times faster than FedAvg, yet achieving more than 50\% improvement in the classification accuracy with respect to the MNIST dataset.

4: %\end{abstract}

5: