A Stochastic Computational Graph with Ensemble Learning Model for solving Controller Placement Problem in Software-Defined Wide Area Networks

Abstract

The Preponderance of literature has established that most of the metaheuristic algorithms were associated with identified challenges in solving the Controller Placement Problem in SD-WAN. This study proposed a Stochastic Computational Graph Model with an Ensemble Learning (SCGMEL) approach to address the scalability, intelligence, and high computational complexity challenges experienced by the existing metaheuristic algorithms. The proposed SCGMEL used stochastic gradient descent with momentum and learning rate decay, a computational graph model, and the eXtreme Gradient Boosted Trees (XGBoost) algorithm as the optimization and machine learning approaches. The proposed solution was tested using datasets from Internet Zoo topology with six objective functions: load balancing, maximum controller failure, average controller-to-controller latency, average switch-to-controller latency, and maximum controller-to-controller latency. The XGBoost outperformed other regression models, in predicting the number of controllers, with mean absolute error of 1.855751 versus 1.883536, 3.729863, and 3.829268 for the random forest, logistic regression, and K-nearest neighbor, respectively. Furthermore, the execution time, average and total CPU usages of the algorithms demonstrated the computational efficiency of the proposed SCGMEL over ANSGA-III, NSGA-II, and MOPSO with percentage decreases of 99.983%, 99.985%, and 99.446%, respectively. Consequently, the proposed SCGMEL was recommended for controller placement in SD-WAN, subject to the usage conditions.