Learning to Dynamically Allocate Radio Resources in Mobile 6G in-X Subnetworks

This paper investigates efficient deep learning based methods for interference mitigation in independent wireless subnetworks via dynamic allocation of radio resources. Resource allocation is cast as a mapping from interference power measurements at each subnetwork to a class of shared frequency channels. A deep neural network (DNN) is then trained to approximate this mapping using data obtained via application of centralized graph coloring (CGC). The trained network is then deployed at each subnetwork for distributed channel selection. Simulation results in an environment with mobile subnetworks have shown that relatively small-sized DNNs can be trained offline to perform distributed channel allocation. The results also show that regardless of the choice of initialization, a DNN for distributed channel selection can achieve similar performance as CGC up to a probability of loop failure (PLF) of 6 x 10(-5) in diverse environments with only aggregate interference power measurements as input.