Shared-resource Generative Adversarial Network (GAN) training for 5g Ultra Reliable Low Latency Communication (URLLC) deep reinforcement learning augmentation

Abstract

Recently, Deep Reinforcement Learning (DRL) has demonstrated exceptional abilities in finding high-quality real-time solutions. Nonetheless, DRL encounters difficulties in managing rare events in wireless communications, primarily because DRL agents lack experience with such situations during their training. This challenge is particularly crucial for 5G ultra-reliable lowlatency communications (URLLC), where rare events can significantly reduce communication reliability. To address this issue, recent efforts have focused on utilizing data augmentation techniques, such as Generative Adversarial Networks (GANs), to enable DRL agents to learn to handle rare events by exposing them to a diverse range of scenarios. However, training GANs in resourceconstrained 5G networks, like Ultra Dense Networks (UDNs), remains a challenging task due to the substantial computational power required, especially when processing high-dimensional data. Despite its importance, there has been limited attention on effectively training GANs within resource-limited environments. In this thesis, I introduce a novel architecture and an optimization problem for training GANs in resource-constrained 5G networks. My proposed architecture efficiently shares limited computing and bandwidth resources between the cloud and edge environments for training multiple GANs. Then, the augmented datasets from the trained GANs are combined with real datasets to be used for training DRL models. I present a mathematical model called OGAN, designed to optimize the allocation of computation and communication resources for GAN training, aiming to enhance DRL reliability, a critical aspect of URLLC. Since OGAN is a challenging mixed-integer non-convex problem, I approximate it using the Difference of Convex Functions programming approach and solve it with a Convex-Concave Algorithm. Finally, I carry out extensive simulations to evaluate OGAN’s performance in comparison with two baseline methods.

Date	27 Aug 2024
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Kim Khoa Nguyen (Supervisor)