Wireless Resource Allocation via DNN Permutation Matrix Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing resource allocation methods in wireless communication systems face high computational complexity and inefficiencies, particularly in nonlinear objective functions, and gradient-based training algorithms struggle with gradient loss and require post-processing, limiting optimal allocation performance.
Innovation Solution
A deep learning-based method using a binary variable learning approach and a modified Sinkhorn operation to convert non-convex problems into solvable forms, enabling efficient resource allocation through a neural network with a new activation function, allowing direct derivation of a permutation matrix without gradient loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the Hungarian algorithm is used to solve linear sum allocation problems, then optimal performance is guaranteed, but computational complexity becomes excessively high due to repetitive calculations
Solution Approach 1:
The patent replaces the traditional Hungarian algorithm (a mechanical iterative optimization process) with a deep neural network that learns optimal allocation patterns through training. The DNN substitutes the step-by-step mechanical calculation with a trained model that directly outputs near-optimal solutions, dramatically reducing computational complexity while maintaining allocation optimality.
Solution Approach 2:
The patent performs preliminary training of the deep neural network offline using allocation problem data and optimal solutions from the Hungarian algorithm. This preliminary learning phase allows the system to store learned patterns and heuristics, enabling fast inference during actual resource allocation without repeating complex optimization calculations in real-time.
2Reliability
If a solver for mixed integer programs is used to solve allocation problems with convex objective functions, then feasible solutions are obtained, but computational complexity increases due to repetitive calculation
Solution Approach 1:
The patent replaces the mixed integer program solver (a complex iterative mathematical optimization system) with a deep neural network that has been trained to directly predict feasible and near-optimal allocation solutions. The DNN substitution eliminates the need for repetitive constraint checking and iterative optimization, maintaining solution feasibility while reducing computational burden.
3Productivity
If supervised learning is used to memorize solutions from existing optimization techniques, then performance close to optimal is achieved with reduced execution time, but additional data collection and training processes are required
Solution Approach 1:
The patent performs the data collection and model training as a preliminary offline action. During the training phase, the system collects allocation problems and their optimal solutions, trains the deep neural network to learn the mapping, and stores the learned parameters. Once trained, the model executes rapidly during inference without requiring additional data collection or training processes, achieving both high speed and simplicity.
4Device complexity
If existing deep learning structures are used for resource allocation, then reduced complexity is achieved, but post-processing is always required which results in performance loss
Solution Approach 1:
The patent employs a dynamic deep neural network architecture that adaptively adjusts its output to directly produce valid permutation matrices representing resource allocation solutions. The network is designed with appropriate activation functions and output layers that inherently satisfy the constraints of the allocation problem, eliminating the need for post-processing while maintaining both low complexity and high performance.
Data Source
AI summary
The present disclosure relates to a machine learning-based resource allocation method and device in a wireless communication system. A method of performing resource allocation in a wireless communication system according to an embodiment of the present disclosure may include acquiring DNN output information through a deep neural network (DNN) based on input H corresponding to wireless channel-related information; acquiring resource allocation matrix X having a form of a permutation matrix through an approximation operation to a doubly stochastic matrix based on input A corresponding to the DNN output information; and allocating M resource elements belonging to a set of first type resources to N resource elements belonging to a set of second type resources based on the resource allocation matrix.