Operador de recombinação para programação genética baseado em regressão linear múltipla

Abstract

Genetic Programming (GP) is a technique of Evolutionary Computation that evolves individuals of variable size and shape. The expressiveness in the representation makes GP a very useful tool in Engineering, producing competitive results with human intelligence. One of its most common applications is the automatic discovery of models from data analysis, which is known as Symbolic regression. The use of GP requires careful implementation of its genetic operators, especially recombination and mutation. Classical approaches to operator creation are often based on the syntactic characteristics of individuals, while recent techniques are directly or indirectly guided by semantics. There are also approaches that combine Regression Analysis and Evolutionary Computation for higher quality responses. In general, all these operators are subject to the same problem: the number of genes that make up individuals begins to increase wildly after a few generations without improvements in aptitude, compromising the quality of the responses produced. This phenomenon is called bloat effect. The objective of this work is to present a new operator for Genetic Programming that allows the evolution of populations that can present more accurate and structural individuals whose size is naturally controlled. The developed operator simultaneously acts on recombination and mutation, promoting variational inheritance and population diversity. By providing the production of high-quality individuals while evolving populations without the harmful effects associated with bloat, the developed operator proved to be superior to classical subtree recombination, to new genetic operators based on semantics, and also to other recent techniques based on Analysis of Regression.