Federated Learning Based Resource Allocation for Wireless Communication Networks

In this paper we introduce federated learning (FL) based resource allocation (RA) for wireless communication networks, where users cooperatively train a RA policy in a distributed scenario. The RA policy for each user is represented by a local deep neural network (DNN), which has the same structure for all users. Each DNN monitors local measurements and outputs a power allocation to the user. The proposed approach is model-free; each user is responsible for training its own DNN to maximize the sum rate (SR) and communicates with the server to aggregate its local DNN with other DNNs. More importantly, each user needs to probe only its own data rate as a distributed reward function and communications with the server once in a while. Simulations show that the proposed approach enables conventional deep learning (DL) based RA methods to not only use their policy in a distributed scenario, but also to (re)train their policy in time-varying environments in a model-free distributed manner without needing a computationally complex server.