Decentralized Learning Framework for Hierarchical Wireless Networks: A Tree Neural Network Approach

Abstract

This article presents a flexible deep-learning strategy that tackles decentralized optimization tasks in multitier networks where wireless nodes are deployed in a hierarchical structure. Practical multitier networks have arbitrary node populations as well as their backhaul connections. Thus, node operations in the multitier network request versatile inference rules for arbitrary network configurations. To this end, we present a tree-based learning strategy which transforms the multitier network optimization into a collaborative inference process over random trees. For the decentralized structure, each node in a tree is equipped with dedicated deep neural network (DNN) modules. A group of these component DNNs builds a tree deep neural network (TNN) where forward pass calculations define the node interaction policy. The TNN is carefully designed such that it can be universally applied to random trees. The training mechanism is developed to involve a number of random tree instances so that the TNN can be generalized to arbitrary network configurations. As a consequence, the TNN can scale up with a large number of nodes which requires only a single training process. The scalability of the proposed framework is validated for various multitier network optimization problems. Numerical results demonstrate the effectiveness of the TNN over existing approaches.