Banca de DEFESA: Renan Santos Mendes

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : Renan Santos Mendes
DATE: 28/03/2022
TIME: 09:00
LOCAL: Videoconferência
TITLE:

**

KEY WORDS:

Vehicle Routing Problem with Demand Responsive Transport. Many Objective Optimization. Dimensionality Reduction Techniques. Cluster Analysis. Feature Selection. Pearson correlation. Kendall’s τ. Maximum Variance. Dispersion Rate. Chord Diagram.

PAGES: 110
BIG AREA: Outra
AREA: Multidisciplinar
SUMMARY:

In this work, a many objective formulation for the Vehicle Routing Problem with Demand Responsive Transport (VRPDRT) is addressed. The problem can be considered as a mode of transport similar to on-demand transportation services with shared races in which passengers are transported from their origin to their destination, sharing the same vehicle. The purpose of this type of transport is to reduce operating/displacement costs, meeting the passengers needs and offering high quality services. This work proposes the use of online cluster analysis based on Pearson’s and Kendall’s τ correlation coefficients to perform the dimensionality reduction at each generation of the MOEA/D algorithm, called OnCL_τMOEA/D and OnCL_ρ-MOEA/D. The first sequence of experiments compared the following approaches: (i) offline clustering using Pearson’s correlation coefficient; (ii) offline clustering using Kendall’s τ correlation; (iii) online clustering using the Pearson correlation coefficient; (iv) online clustering using Kendall’s correlation coefficient τ; and (v) a baseline version, the MOEA/D in the original version. The algorithms were tested using a real dataset containing distances and travel times for Belo Horizonte city. To assess the dispersion of the solutions obtained by the algorithms, the hypervolume indicator was applied. The results of these experiments showed (i) there is no difference in the formulation obtained for the offline approaches; (ii) the online cluster-based algorithms are statistically better than the offline cluster-based algorithm; (iii) it is not possible to state that there is a statistical difference between the algorithms with online dimensionality reduction and the MOEA/D when the hypervolume indicator is used to evaluate the obtained solutions. Chord diagrams were applied to the obtained solutions and indicated that the diversity of solutions obtained by OnCL_ρ-MOEA/D and OnCL_τ-MOEA/D are also slightly different. The second sequence of experiments performed compares the dimensionality reduction approaches based on online aggregation and another based on attribute selection. The approach using feature selection eliminates the objectives using two unsupervised feature selection techniques: Dispersion Rate (DR) and Maximum Variance (MV). Feature selection is used to select the cluster representative. The frequency of impact of dimensionality reduction on algorithm performance was also analyzed. Approaches using feature selection are coupled to the MOEA/D algorithm. Three algorithms were tested (i) online cluster using Pearson correlation coefficient with MOEA/D (OnCLρ_g-MOEA/D); (ii) online cluster using Pearson’s correlation coefficient and dispersion rate with MOEA/D (OnDRρ_g-MOEA/D); (iii) online cluster using Pearson’s coefficient and Maximum Variance with MOEA/D (OnMVρ_g-MOEA/D), where g = 1,25,50,100. The results of these new tests showed that, regardless of the frequency of the dimensionality reduction approach and the instances, the algorithms MOEA/D and OnCLρ_g-MOEA/D are statistically superior to OnDRρ_g-MOEA/D and OnMVρ_g-MOEA/D. It is not possible to state that there is a statistical difference between the results of the algorithms based on the selection of attributes online in terms of the hypervolume indicator. The chord diagrams showed that there is a greater diversity of solutions obtained when all objective functions are used in the reduced form of the problem and also when the algorithms remain for more generations in a given reduced formulation.

BANKING MEMBERS:
Interno - ALISSON MARQUES DA SILVA
Externa à Instituição - CAROLINA GIL MARCELINO - UFRJ
Presidente - ELIZABETH FIALHO WANNER
Interno - FLAVIO VINICIUS CRUZEIRO MARTINS
Interno - GUSTAVO CAMPOS MENEZES
Externo à Instituição - IVAN REINALDO MENEGHINI - IFMG
Externo ao Programa - JOAO FERNANDO MACHRY SARUBBI