- author = "Bing Li",
- title = "Study on Genetic Network Programming with Variable Size Structure and Genotype/Phenotype Mapping Mechanism",
- school = "Graduate School of Information, Production and Systems, Waseda University",
- year = "2013",
- address = "Japan",
- month = jun,
- keywords = "genetic algorithms, genetic programming, GNP, Genotype/Phenotype",
-
URL = "
http://hdl.handle.net/2065/44360",
-
URL = "https://waseda.repo.nii.ac.jp/?action=repository_uri&item_id=16739&file_id=20&file_no=3",
-
URL = "
http://hdl.handle.net/2065/44360/Honbun-6415.pdf",
-
URL = "https://dl.ndl.go.jp/info:ndljp/pid/8977235",
- size = "124 pages",
-
abstract = "In the research field of Artificial Intelligence,
Evolutionary Algorithms (EAs) are subset of important
optimization technologies, which are inspired by the
Darwin theory of evolution. EAs are kinds of effective
algorithms for solving very large search space problems
with less prior knowledge and human intervention.
Starting from the 1950s, a lot of EAs are developed,
such as Evolution Strategies (ES), Genetic Algorithm
(GA), Genetic Programming (GP) and Evolution
Programming (EP). They have been successfully applied
to many fields such as engineering, biology, economics,
marketing, robotics, physics, chemistry, education and
so on.
After investigating the benefits and shortcoming of GA and GP, Genetic Network Programming (GNP) was proposed around 2000. The directed graph structure of GNP extends the chromosome representation of strings in GA and trees in GP, which makes it have high expression ability with relevant small size of individuals, and consequently GNP has the better performance than other evolutionary algorithms. Nowadays, GNP is not only used to solve benchmark problems but also applied to many real world applications such as elevator supervisory control systems, stock market prediction, datamining and traffic prediction.
Since GNP was proposed, many methods have been developed to improve the performance of GNP such as combining GNP with reinforcement learning, introducing symbiotic learning in GNP, upgrading the structure of GNP by defining macro node and rule accumulation. Although these methods have been proved to improve the performance of GNP by combing some other machine learning methods, some useful prior knowledge of biology: variable length of gene, evolution by gene duplication and genotype-phenotype mapping, are not well considered. Therefore, in this research, two kinds of methods and their extensions have been proposed to improve the performance including the expression and generalization ability of GNP by upgrading the structure of GNP using the above theories, and to solve two problems of GNP, i.e., the node size of GNP is fixed and an individual is a solution.
One of the methods is Variable Size Genetic Network Program (GNPvs) and its extension GNPvs with Replacement (GNPvs-R), which simulates the variable length of gene and gene duplication, and solve the problem that the node size of GNP is fixed. The other is Genetic Network Programming for Automatic Program Generation with Mapping Mechanism (GNP-APGm) and its extension Subroutine embedded GNP-APGm (GNPsr-APGm), which implements the genotype-phenotype mapping process, and solve the problem that an individual is a solution.
The above methods are verified on the tileworld benchmark problem. The simulation results shows these proposed methods increase the performance of GNP exactly.",
-
notes = "info:ndljp/pid/8977235
supervisor: Hirasawa",
