Development of a linear genetic programming algorithm using an estimation of distribution algorithm applied to supervised machine learning

• Grant number: 18/13202-4
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate (Direct)
• Start date: September 03, 2018
• End date: September 02, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Márcio Porto Basgalupp
• Grantee: Léo Françoso Dal Piccol Sotto
• Supervisor: Franz Rothlauf
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil
• Institution abroad: Johannes Gutenberg University Mainz (JGU), Germany
• Associated to the scholarship: 16/07095-5 - Development of the probabilistic linear genetic programming technique and application on Kaizen programming for supervised machine learning, BP.DD

Abstract

Linear Genetic Programming (LGP) is a kind of algorithm capable of evolving computer code which has been successfully applied in various problems, such as Machine Learning (ML), navigation, and routing. As in other Evolutionary Algorithms (EAs), its stochastic search process neither has the knowledge to produce good quality solutions nor is able to avoid poor quality solutions, which reduces its efficacy. Furthermore, their recombination operators often ignore the correlation between the different positions in a genotype. In order to deal with this problem in EAs, researchers proposed the Estimation of Distribution Algorithm (EDA) that uses a probability model, built from good quality solutions, to sample promising solutions, which in turn are used to update the model. In order to lessen the issue with the disruptive recombination operators, bivariate and multivariate models also explore the correlation between the different string positions. The PhD project associated with this proposal has the objective of studying and proposing different probability models for LGP: univariate, bivariate, and multivariate. The resulting Probabilistic Linear Genetic Programming (PLGP) techniques will be applied on supervised ML problems, such as regression and classification. During the research stay period, the student will develop a multivariate model for LGP. The original aspect of the work is either the adaptation of existing multivariate models or the implementation of new models specifically suited to the linear structure of LGP, while its main relevance is the application on supervised ML problems. (AU)

Scientific publications

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

DAL PICCOL SOTTO, LEO FRANCOSO; ROTHLAUF, FRANZ; LOPEZIBANEZ, M. On the Role of Non-effective Code in Linear Genetic Programming. PROCEEDINGS OF THE 2019 GENETIC AND EVOLUTIONARY COMPUTATION CONFERENCE (GECCO'19), v. N/A, p. 9-pg., 2019-01-01. (18/13202-4)

DE MELO, VINICIUS V.; SOTTO, LEO F. D. P.; LEONARDO, MATHEUS M.; FARIA, FABIO A.. Automatic Meta-Feature Engineering for CNN Fusion in Aerial Scene Classification Task. IEEE Geoscience and Remote Sensing Letters, v. 17, n. 9, p. 1652-1656, SEPT 2020. (18/13202-4, 16/07095-5)

SOTTO, LEO FRANCOSO DAL PICCOL; ROTHLAUF, FRANZ; DE MELO, VINICIUS VELOSO; BASGALUPP, MARCIO P.. An Analysis of the Influence of Noneffective Instructions in Linear Genetic Programming. EVOLUTIONARY COMPUTATION, v. 30, n. 1, p. 24-pg., 2022-03-01. (18/13202-4, 16/07095-5)