SIS PUT | Semantic Extensions for Genetic Programming

Scientific Information System of the Poznań University of Technology

PL EN

Main page / Dissertations / Semantic Extensions for Genetic Programming

Submit a comment

Dissertation

Download BibTeX

Title

Semantic Extensions for Genetic Programming

Authors

Bartosz Wieloch (WI) ^{[ 1 ][ P ]}

^{[ 1 ]} Instytut Informatyki, Wydział Informatyki, Politechnika Poznańska | ^{[ P ]} employee

Promoter

Krzysztof Krawiec (WI) ^{[ 1 ][ P ]}

^{[ 1 ]} Instytut Informatyki, Wydział Informatyki, Politechnika Poznańska | ^{[ P ]} employee

Reviewers

Title variant

PL Semantyczne rozszerzenia programowania genetycznego

Language

english

Abstract

EN Genetic Programming (GP) is a most popular approach to automatic generation of computer programs. Standard methods applied in GP use raw fragments of evolved programs to construct new, hopefully better ones. These methods, except the selection phase, pass over the behavior of the modified programs and operate mainly on their syntax. In this dissertation we follow alternative, semantic-oriented approach that concentrates on the actual behavior of programs in population to determine how to construct the new ones. This research trend grew up as an attempt to overcome weaknesses of methods that rely only on syntax analysis. Recent contributions suggest that semantic extensions to GP can be a remedy to poor performance of classical, syntactic methods. Therefore, in this dissertation we firstly present the advantages and disadvantages of several possible descriptions of program’s behavior. Then we introduce the concept of semantics used in all semantic extensions presented throughout this thesis. The first semantic extension presented in this thesis is a method population initialization which forces the individuals in population to be semantically unique. We also show selected semantic-aware variants of crossover and mutation operators. In particular, we test how they perform with and without our initialization method. Next, we introduce and formalize our novel concept of desired semantics that describes the desired behavior for given part of a program. Then we propose several methods that employ desired semantics to create new programs by combining matching parts. We show that some of these methods significantly outperform other methods, semantic as well as syntactic ones. The second important proposition of this thesis is the concept of functional modularity. Functional modularity involves defining modules based on their semantic properties instead of syntactical ones, like, e.g., the frequency of occurring some code fragments. Functional modularity can be used to decompose a problem into potentially easier parts (subproblems), and then to solve the subproblems in isolation or together. All the described methods are illustrated with extensive experimental verification of their performance on a carefully prepared benchmark suite that contains problems from various domains. On this suite, we show the overall advantage of semantic-aware extensions, especially for methods that rely on desired semantics.

Number of pages

159

Signature of printed version

DrOIN 1496

On-line catalog

to201380290

Comments

Przewód doktorski na starych zasadach – brak niektórych informacji

Full text of dissertation

no permission to download file

Access level to full text