The Multi-Objective Evolutionary Algorithm based on Decomposition with Dynamical Resource Allocat... more The Multi-Objective Evolutionary Algorithm based on Decomposition with Dynamical Resource Allocation (MOEA/D-DRA) has obtained very good results on various multi-objective optimization problems in the past few years. This paper focuses on an attempt to improve even more its performance by introducing a hyper-heuristic mechanism to select the best set of its operators and parameters. In this paper we use Upper Confidence Bound (UCB) as the basis of the hyper-heuristic, and test three versions of the proposed approach. Four well known benchmarks (CEC 2009, WFG, DTLZ and ZDT) and a quality indicator (hypervolume) are used to analyze the performance of the three variants. The proposed approach is compared with the original MOEA/D-DRA and the results show that tuning the parameters via UCB is an interesting alternative for a hyper-heuristic based version of MOEA/D-DRA on the addressed problems.
2016 5th Brazilian Conference on Intelligent Systems (BRACIS), 2016
Many-objective optimization (four or more objectives) presents many challenges to be considered, ... more Many-objective optimization (four or more objectives) presents many challenges to be considered, highlighting the need to create better algorithms prepared to deal efficiently with the increasing number of objectives. One such challenge is to determine the most efficient operator or combination of operators to be used during the optimization. In order to deal with this challenge, we propose the use of adaptive operator selection mechanisms in many-objective optimization algorithms. Two adaptive operator selection mechanisms, Adaptive Pursuit (AP) and Probability Matching (PM), are incorporated into the NSGA-III framework (a recently proposed, state-of-the-art algorithm to solve many-objective problems) to autonomously select the most suitable operator while solving a many-objective problem, according to the previous performance of each operator. The proposed algorithms, NSGA-IIIAP and NSGA-IIIPM, are tested in four different multi-objective problems from the DTLZ test suite with 3 up to 20 objectives. Statistical tests were performed to infer the significance of the results. The hypothesis that adaptive ways to select the operator to be applied during each stage of the evolutionary process is an effective way to improve the performance of the NSGA-III framework is corroborated by our results.
Proceedings of the Genetic and Evolutionary Computation Conference Companion
The simultaneous optimization of multiple objectives arises in several problems in different disc... more The simultaneous optimization of multiple objectives arises in several problems in different disciplines. This optimization, mainly for many-objective problems brings challenges to the state-of-the-art Multi-Objective Evolutionary Algorithms. Given the various characteristics of the different problem instances and also the features of the algorithms, no single algorithm performs well in all problem instances. Although, if the algorithms characteristics could be combined, cooperatively, to face the problem together, the search ability can be improved. In this work, we evaluate this research question and propose a distributed framework for cooperation of Many-objective Evolutionary Algorithms. In the framework, different algorithms can be executed simultaneously, with small sub-populations and collectively solve the problem instance. The framework performs the cooperation by sharing solutions between the subpopulations. In this way, sharing the information learned from one algorithm to the other. The framework is evaluated using two state-of-the-art algorithms for cooperation, and compared to the algorithms executed alone. The results indicate that the cooperation improves the convergence and diversity of the algorithms in most problem instances. The obtained results motivate the investigation of future works on the proposed framework.
2018 7th Brazilian Conference on Intelligent Systems (BRACIS), 2018
Meta-heuristic algorithms have been used to obtain feasible solutions in reasonable time for many... more Meta-heuristic algorithms have been used to obtain feasible solutions in reasonable time for many NP-hard search problems. However, the performance of the algorithms heavily depends on the features of the problem. So, techniques for automatically choosing or designing algorithms have received attention from researchers in the past years. In this research, we investigate automated algorithm selection for the Quadratic Assignment Problem using a multi-label meta-learning approach. In multi-label learning, an instance can have more than one label at the same time, which is an adequate configuration for this scenario, since there can be ties in the performances of the algorithms. Two approaches for dealing with multi-label learning are compared, regarding both the prediction of which algorithms are the best and the resulting costs of the solutions. The experiments show that the algorithm selection process can yield better results than if we choose only one single algorithm to solve all ...
ADCAIJ: ADVANCES IN DISTRIBUTED COMPUTING AND ARTIFICIAL INTELLIGENCE JOURNAL, 2015
The Set Covering Problem (SCP) is a NP-hard combinatorial optimization problem that is challengin... more The Set Covering Problem (SCP) is a NP-hard combinatorial optimization problem that is challenging for meta-heuristic algorithms. In the optimization literature, several approaches using meta-heuristics have been developed to tackle the SCP and the quality of the results provided by these approaches highly depends on customized operators that demands high effort from researchers and practitioners. In order to alleviate the complexity of designing metaheuristics, a methodology called hyper-heuristic has emerged as a possible solution. A hyper-heuristic is capable of dynamically selecting simple low-level heuristics accordingly to their performance, alleviating the design complexity of the problem solver and obtaining satisfactory results at the same time. In a previous study, we proposed a hyper-heuristic approach based on Ant Colony Optimization (ACO-HH) for solving the SCP. This paper extends our previous efforts, presenting better results and a deeper analysis of ACO-HH parameters...
23rd International Conference of the Chilean Computer Science Society, 2003. SCCC 2003. Proceedings.
Prediction models are fundamental in the early stages of the software development when many times... more Prediction models are fundamental in the early stages of the software development when many times, decisions must be taken without the required information. A typical information that is not available in these stages is software size metrics, such as lines of code (LOC). Models for LOC estimation are obtained from historical data and statistical regression methods are usually applied. These
A análise de processos judiciais é uma tarefa cara, que requer muito tempo de juizes e assessores... more A análise de processos judiciais é uma tarefa cara, que requer muito tempo de juizes e assessores, seja para tomar decisões, seja para classificar de acordo com a jurisprudência vigente. Porém, esse processo é repetitivo e extrair a semântica desse corpus pode ser uma etapa de apoio a esse processo. O objetivo desta pesquisa é desenvolver uma metodologia capaz de gerar automaticamente classificações de documentos jurídicos, utilizando técnicas de processamento de linguagem natural. Primeiramente, coletamos 430.000 sentenças de tribunais trabalhistas brasileiros de 2006 a 2018. Então propomos o uso de técnicas de geração de representação de palavras para representação de dados. Em seguida, usamos técnicas de agrupamento para agrupar semanticamente as decisões judiciais semelhantes. Finalmente, os grupos são usados para criar rótulos artificiais para cada documento. Por fim, utilizamos técnicas de classificação para produzir modelos capazes de captar a semântica do texto judicial. O...
The Multi-Objective Evolutionary Algorithm based on Decomposition with Dynamical Resource Allocat... more The Multi-Objective Evolutionary Algorithm based on Decomposition with Dynamical Resource Allocation (MOEA/D-DRA) has obtained very good results on various multi-objective optimization problems in the past few years. This paper focuses on an attempt to improve even more its performance by introducing a hyper-heuristic mechanism to select the best set of its operators and parameters. In this paper we use Upper Confidence Bound (UCB) as the basis of the hyper-heuristic, and test three versions of the proposed approach. Four well known benchmarks (CEC 2009, WFG, DTLZ and ZDT) and a quality indicator (hypervolume) are used to analyze the performance of the three variants. The proposed approach is compared with the original MOEA/D-DRA and the results show that tuning the parameters via UCB is an interesting alternative for a hyper-heuristic based version of MOEA/D-DRA on the addressed problems.
2016 5th Brazilian Conference on Intelligent Systems (BRACIS), 2016
Many-objective optimization (four or more objectives) presents many challenges to be considered, ... more Many-objective optimization (four or more objectives) presents many challenges to be considered, highlighting the need to create better algorithms prepared to deal efficiently with the increasing number of objectives. One such challenge is to determine the most efficient operator or combination of operators to be used during the optimization. In order to deal with this challenge, we propose the use of adaptive operator selection mechanisms in many-objective optimization algorithms. Two adaptive operator selection mechanisms, Adaptive Pursuit (AP) and Probability Matching (PM), are incorporated into the NSGA-III framework (a recently proposed, state-of-the-art algorithm to solve many-objective problems) to autonomously select the most suitable operator while solving a many-objective problem, according to the previous performance of each operator. The proposed algorithms, NSGA-IIIAP and NSGA-IIIPM, are tested in four different multi-objective problems from the DTLZ test suite with 3 up to 20 objectives. Statistical tests were performed to infer the significance of the results. The hypothesis that adaptive ways to select the operator to be applied during each stage of the evolutionary process is an effective way to improve the performance of the NSGA-III framework is corroborated by our results.
Proceedings of the Genetic and Evolutionary Computation Conference Companion
The simultaneous optimization of multiple objectives arises in several problems in different disc... more The simultaneous optimization of multiple objectives arises in several problems in different disciplines. This optimization, mainly for many-objective problems brings challenges to the state-of-the-art Multi-Objective Evolutionary Algorithms. Given the various characteristics of the different problem instances and also the features of the algorithms, no single algorithm performs well in all problem instances. Although, if the algorithms characteristics could be combined, cooperatively, to face the problem together, the search ability can be improved. In this work, we evaluate this research question and propose a distributed framework for cooperation of Many-objective Evolutionary Algorithms. In the framework, different algorithms can be executed simultaneously, with small sub-populations and collectively solve the problem instance. The framework performs the cooperation by sharing solutions between the subpopulations. In this way, sharing the information learned from one algorithm to the other. The framework is evaluated using two state-of-the-art algorithms for cooperation, and compared to the algorithms executed alone. The results indicate that the cooperation improves the convergence and diversity of the algorithms in most problem instances. The obtained results motivate the investigation of future works on the proposed framework.
2018 7th Brazilian Conference on Intelligent Systems (BRACIS), 2018
Meta-heuristic algorithms have been used to obtain feasible solutions in reasonable time for many... more Meta-heuristic algorithms have been used to obtain feasible solutions in reasonable time for many NP-hard search problems. However, the performance of the algorithms heavily depends on the features of the problem. So, techniques for automatically choosing or designing algorithms have received attention from researchers in the past years. In this research, we investigate automated algorithm selection for the Quadratic Assignment Problem using a multi-label meta-learning approach. In multi-label learning, an instance can have more than one label at the same time, which is an adequate configuration for this scenario, since there can be ties in the performances of the algorithms. Two approaches for dealing with multi-label learning are compared, regarding both the prediction of which algorithms are the best and the resulting costs of the solutions. The experiments show that the algorithm selection process can yield better results than if we choose only one single algorithm to solve all ...
ADCAIJ: ADVANCES IN DISTRIBUTED COMPUTING AND ARTIFICIAL INTELLIGENCE JOURNAL, 2015
The Set Covering Problem (SCP) is a NP-hard combinatorial optimization problem that is challengin... more The Set Covering Problem (SCP) is a NP-hard combinatorial optimization problem that is challenging for meta-heuristic algorithms. In the optimization literature, several approaches using meta-heuristics have been developed to tackle the SCP and the quality of the results provided by these approaches highly depends on customized operators that demands high effort from researchers and practitioners. In order to alleviate the complexity of designing metaheuristics, a methodology called hyper-heuristic has emerged as a possible solution. A hyper-heuristic is capable of dynamically selecting simple low-level heuristics accordingly to their performance, alleviating the design complexity of the problem solver and obtaining satisfactory results at the same time. In a previous study, we proposed a hyper-heuristic approach based on Ant Colony Optimization (ACO-HH) for solving the SCP. This paper extends our previous efforts, presenting better results and a deeper analysis of ACO-HH parameters...
23rd International Conference of the Chilean Computer Science Society, 2003. SCCC 2003. Proceedings.
Prediction models are fundamental in the early stages of the software development when many times... more Prediction models are fundamental in the early stages of the software development when many times, decisions must be taken without the required information. A typical information that is not available in these stages is software size metrics, such as lines of code (LOC). Models for LOC estimation are obtained from historical data and statistical regression methods are usually applied. These
A análise de processos judiciais é uma tarefa cara, que requer muito tempo de juizes e assessores... more A análise de processos judiciais é uma tarefa cara, que requer muito tempo de juizes e assessores, seja para tomar decisões, seja para classificar de acordo com a jurisprudência vigente. Porém, esse processo é repetitivo e extrair a semântica desse corpus pode ser uma etapa de apoio a esse processo. O objetivo desta pesquisa é desenvolver uma metodologia capaz de gerar automaticamente classificações de documentos jurídicos, utilizando técnicas de processamento de linguagem natural. Primeiramente, coletamos 430.000 sentenças de tribunais trabalhistas brasileiros de 2006 a 2018. Então propomos o uso de técnicas de geração de representação de palavras para representação de dados. Em seguida, usamos técnicas de agrupamento para agrupar semanticamente as decisões judiciais semelhantes. Finalmente, os grupos são usados para criar rótulos artificiais para cada documento. Por fim, utilizamos técnicas de classificação para produzir modelos capazes de captar a semântica do texto judicial. O...
Uploads
Papers by Aurora Pozo