Regularization Strategy Aided Robust Unsupervised Learning for Wireless Resource Allocation

Abstract

Unsupervised learning (UL) is widely used in the wireless resource allocation problems due to its lower computational complexity and better performance compared with traditional optimization algorithms. Since wireless resource allocation problems usually have several constraints, primal-dual learning based UL framework are widely adopted. However, the primal-dual learning approach has the problem of oscillation around the constraint threshold while training and there may be serious constraint violations when deployment. In addition, although the output of the neural network can also be restricted to the feasible region by the penalty function method, the optimality of such training methods cannot be guaranteed. In this paper, we combine the primal dual learning method with the penalty function method and propose a regularized unsupervised learning (RUL) framework to enhance the robustness of the primal-dual learning based UL framework. In the proposed RUL framework, we use regularization techniques to improve the robustness of primal-dual learning by reducing the risk of constraint violations while training. A quadratic penalty term is introduced into the Lagrangian function of the wireless optimization problem where the constraints can be equivalent to equality constraints to form its augmented Lagrangian function. In the simulation, we give a simple point to point power optimization problem as an example to show that the proposed RUL can improve the robustness of constraint convergence, and can also accelerate training speed.