FL-ORA: Optimized and Decentralized Resource Allocation Scheme for D2D Communication

Abstract

This article presents an optimized and decentralized resource allocation approach aimed at maximizing the system throughput and energy efficiency of device-to-device (D2D) communication. The proposed scheme modifies the meta-heuristic whale optimization algorithm (WOA) by blending the differential evolution (DE) technique in the WOA’s exploration phase to offer intelligence and reduce the computational overburden. The hybrid WOA (DE+WOA) serves as a physical layer access control that efficiently finds the optimal cellular users (CUs) and D2D users (DUs) based on their channel conditions. The proposed access control acts as a restrictive filter, where only optimal CU-DUs can participate in resource allocation tasks. Furthermore, a dataset has been prepared using the optimal CUs-DUs channel conditions from the hybrid WOA to serve as input for the federated learning (FL)-based resource allocation. We utilized statistical tests (e.g., Spearman’s test) to analyze the generated dataset’s non-independent and identically distributed (non-IID) characteristics, thus providing generalization in the AI training. Allowing only the optimal CUs and DUs (from hybrid WOA) in the FL-based resource allocation substantially reduces the computational cost of AI training and improves energy efficiency. In the FL-based resource allocation, we used a sequential convolutional neural network (CNN) trained on the aforementioned dataset to provide proactive resource allocation decisions. Furthermore, we used momentum-based weight aggregation in the FL to reduce the computational burden on the central server. The proposed scheme is assessed by utilizing different standard metrics, such as training accuracy (98.93%), training time, overall system throughput (35.62 Mbps), energy efficiency (96.42 bits/joule), and resource fairness.