Adversarial Training: Enhancing Out-of-Distribution Generalization for Learning Wireless Resource Allocation

Abstract

Deep neural networks (DNNs) have widespread applications for optimizing resource allocation. Yet, their performance is vulnerable to distribution shifts between training and test data, say channels. In this letter, we resort to adversarial training (AT) for enhancing out-of-distribution (OOD) generalizability of DNNs trained in unsupervised manner. We reformulate AT to capture the OOD degradation, and propose a one-step gradient ascent method for AT. The proposed method is validated by optimizing hybrid precoding. Simulation results showcase the enhanced OOD performance of multiple kinds of DNNs across various channel distributions, when only Rayleigh fading channels are used for training.