Evaluation of topology optimization objectives in IP networks

Abstract

In the past, various optimization objective functions have been proposed to help in network optimization, especially for use in traffic engineering (TE) and topology optimization. This variety of optimization objectives resulted in the emergence of algorithms targeting different objectives. However, the role of the objective function has been largely overlooked. Because, the choice of a particular objective function was not justified in most of the cases. Some researchers criticized this arbitrary selection of objective functions. Even though some researchers intuitively suggest using a specific objective, only few work tackled with the problem of evaluating the objectives. In this paper, we evaluate various network optimization objectives on topology optimization. Previously, a study analyzed the efficiency of some routing optimization objectives using linear programming (LP) by linear relaxation. However, some of the objective functions are nonlinear, and such a linear relaxation does not treat each objective equally. The difficulty arises due to the fact that optimization algorithms are objective function tailored heuristics. To achieve fairness, we compare and analyze different traffic optimization objectives for topology optimization using neural networks which are used to model nonlinear relations. By using neural networks, we strive to avoid any unfairness, such as obviating linear approximation. Also, our work suggests which features are meaningful for machine learning in network optimization. Our method partially agrees with the previous work, and we conclude that delay is the best performing optimization objective.