Utilização de técnicas de redução de dimensionalidade em algoritmos de otimização com muitos objetivos no problema de sincronização de semáforos

Abstract

Urban mobility is a current problem of modern society and large urban centers. Intelligent Transportation Systems (ITS) use technology to address these mobility issues. In the context of ITS, traffic management is an area that uses new concepts of organization and maintenance of traffic, seeking to achieve a flow of quality traffic. The traffic light synchronization is one of them and its main objective is to ensure that the vehicles have a good fluidity in the traffic, ensuring to cross a route in the shortest possible time. With timing achieved, quality measures tend to improve, such as reducing pollutant emissions, fuel consumption, delay time, overall average speed, and so on. Indicating the best traffic light is a very complex task. It is difficult to model a real situation because there are chains of crosses with different characteristics. The optimization in semaphore synchronization classifies as NP-Complete problem, the difficulty of the problem grows exponentially, when the numbers of decision variables and measures of qualities increase. Therefore, no classical technique would be able to solve it in a reasonable time. One solution is to model the problem as of optimization, through a traffic simulator. With the simulator, it is able to construct a traffic signal representation, composed of roads, routes, vehicles, intersections and traffic lights. From configurations of flow conditions in different scenarios, we can obtain these measures of qualities, treated as objectives, extracted from the simulator itself. The problem is modeled as multiobjective optimization and by working with more than 3 objective functions, it is classified as optimization with many objectives. Traditional algorithms face problems in optimization with many goals, one of the techniques to solve is the reduction of goals. The objective of this work is to use dimensionality reduction machine learning techniques to reduce objectives in the problem of synchronization of traffic lights. Two techniques were applied in the search to identify the essential objectives and discard the others to reduce. The techniques studied were L-PCA and K-PCA using the polynomial, RBF and sigmoid kernels. An optimization, using the NSGA-II and NSGA-III algorithms, was applied in the sets containing all the objectives, were worked 12, and also for the subsets obtained by the reduction. Comparisons of the optimizations between the full sets and the reduced subsets were made. We also performed tests to identify if there was statistical difference between the algorithms. The results showed that the NSGA-III obtained better results, and the K-PCA with polynomial kernel was the best reduction algorithm, even managing to overcome NSGA-III without reduction. It also concluded that there was no statistical difference between the algorithms, thus, to work with a smaller set of objectives, if it has a better performance in the optimization without losing the information quality.