In this paper, we investigate the computational behavior of the exterior point simplex algorithm. Up until now, there has been a major difference observed between the theoretical worst case complexity and practical performance of simplex-type algorithms. Computational tests have been carried out on randomly generated sparse linear problems and on a small set of benchmark problems. Specifically, 6780 linear problems were randomly generated, in order to formulate a respectable amount of experiments. Our study consists of the measurement of the number of iterations that the exterior point simplex algorithm needs for the solution of the above mentioned problems and benchmark dataset. Our purpose is to formulate representative regression models for these measurements, which would play a significant role for the evaluation of an algorithm's efficiency. For this examination, specific characteristics, such as the number of constraints and variables, the sparsity and bit length, and the condition of matrix A, of each linear problem, were taken into account. What drew our attention was that the formulated model for the randomly generated problems reveal a linear relation among these characteristics.
Selection of the most efficient algorithm for a given set of linear programming problems has been a significant and, at the same time, challenging process for linear programming solvers. The most widely used linear programming algorithms are the primal simplex algorithm, the dual simplex algorithm, and the interior point method. Interested in algorithm selection processes in modern mathematical solvers, we had previously worked on using artificial neural networks to formulate and propose a regression model for the prediction of the execution time of the interior point method on a set of benchmark linear programming problems. Extending our previous work, we are now examining a prediction model using artificial neural networks for the performance of CPLEX’s primal and dual simplex algorithms. Our study shows that, for the examined set of benchmark linear programming problems, a regression model that can accurately predict the execution time of these algorithms could not be formed. Therefore, we are proceeding further with our analysis, treating the problem as a classification one. Instead of attempting to predict exact values for the execution time of primal and dual simplex algorithms, our models estimate classes, expressed as time ranges, under which the execution time of each algorithm is expected to fall. Experimental results show a good performance of the classification models for both primal and dual methods, with the relevant accuracy score reaching 0.83 and 0.84, respectively.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.