复杂约束条件下的无线资源优化:AI视角下的方法和洞察

Wireless Resource Optimization under Complex Constraints:Methods and Insights from an AI Perspective

针对无线资源优化问题中普遍存在的复杂约束条件,详细梳理了各种基于AI的优化方法。尽管现有基于AI的优化方法在各种功率分配和波束赋形设计问题上取得了巨大成功,但大多数方法考虑的问题都仅仅配备较为简单的约束条件(例如功率预算约束条件),这些约束条件可以通过简单的投影操作得到满足。然而,对于更为复杂的约束条件,例如非凸的服务质量约束条件,其优化变量和无线信道通常是耦合的,如何有效处理非凸耦合的约束条件仍然是一个较大挑战。针对复杂约束条件下的无线资源优化问题,将现有基于AI的优化方法分为三类:监督学习方法、惩罚学习方法和拉格朗日对偶方法,并对这三类方法的适用性和不足作了细致的分析。最后,提出了一种基于增广拉格朗日法的惩罚-对偶学习框架,通过交替训练两个独立的神经网络,分别用以推断原问题的解和相应的拉格朗日乘子。此外,将所提出的惩罚-对偶学习框架应用于两个典型的无线资源优化问题上,并通过仿真实验表明,所提出的惩罚-对偶学习框架在约束违反和计算时间方面,分别优于当前最先进的AI和传统优化方法。

This paper summarizes various deep learning-based methods for tackling the complex constraints commonly encountered in wireless resource optimization problems.Although existing deep learning-based methods have achieved great success in various power allocation and beamforming design problems,most of the considered problems are restricted to simple constraints(such as power budget constraints),which can be satisfied by a simple projection operation.However,it is still challenging to tackle the more complex constraints,such as nonconvex quality-of-service constraints,where the optimization variables and wireless channels are highly coupled.Aiming at the wireless resource optimization problem under complex constraints,this paper first introduces the applicability and the deficiencies of existing deep learning-based methods,which are classified into three categories:supervised learning methods,penalty learning methods,and Lagrangian duality methods.Then,a penalty-dual learning framework based on the augmented Lagrangian approach is proposed,which alternately trains two independent neural networks to infer the original variables and the corresponding Lagrange multipliers,respectively.In addition,by applying the proposed penalty-dual learning framework to two typical wireless resource optimization problems,we show through simulation experiments that the proposed penalty-dual learning framework outperforms state-of-the-art deep learning-based methods and traditional optimization algorithms in terms of constraint violation and computational time.