Development of the probabilistic linear genetic programming technique and application on Kaizen programming for supervised machine learning

Abstract

Linear Genetic Programming is a type of algorithm able to evolve code in a programming language. LGP has been applied in a wide range of problems, as machine learning, navigation, and routing. Its stochastic search process has no knowledge to generate good solutions nor is able of avoiding poor ones, which decreases its efficiency. In this project, two topics will be explored in order to tackle these problems: the manipulation of randomness and the composition of solutions from blocks of code (modularity). Regarding the first topic, studies will be made on conditional probability models, originating the Probabilistic Linear Genetic Programming (PLGP). These models will be updated throughout the generations in order to incorporate knowledge about structures that should be used or avoided by individuals. The second topic consists of expanding the study of the Kaizen Programming (KP), which is an approach that composes a complete solution from partial solutions. They can be continually improved by means of an evolutionary technique, like PLGP, that will be developed in this project. Thus, while PLGP can be independently used for solving the same problems as LGP, it will also be employed as a module for the KP, for regression and classification problems, focusing on feature construction. The use of probabilistic models in PLGP, the adding of modularity, and the use with KP are the original aspects of the work, while the most relevant aspect is the application on the automatic construction of high quality feature sets. (AU)