Study on Architecture of Genetic Network Programming based on Cooperative Division Strategy

Created by W.Langdon from gp-bibliography.bib Revision:1.7615

@PhdThesis{YangYang:thesis,

• author = "Yang Yang",
• title = "Study on Architecture of Genetic Network Programming based on Cooperative Division Strategy",
• school = "Waseda University",
• year = "2012",
• address = "Japan",
• month = nov,
• keywords = "genetic algorithms, genetic programming, Genetic Network Programming, coevolution",
• URL = "http://jairo.nii.ac.jp/0069/00023924/en",
• URL = "http://dspace.wul.waseda.ac.jp/dspace/bitstream/2065/40055/1/Honbun-6123.pdf",
• URL = "http://dspace.wul.waseda.ac.jp/dspace/bitstream/2065/40055/2/Shinsa-6123.pdf",
• URL = "http://dspace.wul.waseda.ac.jp/dspace/bitstream/2065/40055/3/Gaiyo-6123.pdf",
• size = "113 pages",
• abstract = "Making computers automatically solve problems is central to Artificial Intelligent and many technologies have been developed under the name of what is called machine learning. Recently developed Genetic Network Programming (GNP) is a graph-based evolutionary algorithm extended from GA and GP. Because GNP represents its solutions using graph structures, which contributes to creating quite compact programs and realising partially observable process, it has extended from purely theoretical concept to real-life applications in a very short time. However, the drive for applying GNP to a wider range of applications should be continued constantly examining the current GNP architectures and making improvements. Many studies have shown that real world problems have hurdles that could not be solved by original GNP. Like many other areas of computer science, GNP can evolve more rapidly and produce better performances with new techniques. The development of advanced techniques to boost GNP performances seems important and promising.

  This thesis studies the following topics related to the architecture and applications of GNP:

  Search space is a key issue while complementing GNP in the area of stock markets. There are many indexes to be considered in the stock markets, moreover their relationships are nonlinear.

  Overfitting generally exists in the machine learning field, which means only specific situations can be handled instead of generalised situations. The possibility of over fitting exists because a model is typically trained using training data by maximising its performances. However, its overall performances are determined not by its performances on the training data but by its ability to perform on unseen data.

  Computational complexity is also needed to be considered, since most machine learning approaches only use a single monolithic system to solve large and complex problems.

  Multitask is a common feature in many real-world problems, but the standard methodology in machine learning is to study them by one system.

  In order to improve the performance of GNP, when faced with the above-mentioned hurdles, this thesis introduces new advanced techniques of GNP. Firstly, hierarchical architecture GNP is proposed, which uses subroutine mechanism, and furthermore the functional subroutines are introduced. Secondly, the division architecture GNP is proposed, which uses cooperative coevolution for the subprograms. Next, the distributed architecture GNP is proposed, which has task program using multitask learning. To illustrate the performances of the proposed methods, two real-world applications are conducted. One is the stock markets, which is one of the targets for most popular investments due to its high expected profits. The second application is the tile-world, where its aim is to find successive optimal behaviours for the multi-tasks making judgements and taking proper actions for the current environments.

  In chapter 1, the research background, objective and outline of the thesis are descried. The objective of this research is to propose the advanced techniques into GNP to develop better methods for real-world applications.",

• abstract = "Chapter 2 proposes a methodology to construct the models for creating trading rules using Genetic Network Programming-Sarsa with Subroutines (GNPsb-Sarsa), which offers an alternative population, where individuals are represented by subroutines. The basic idea of GNP sb-Sarsa is to automatically discover effective subroutines, which capture the underlying structure and building blocks of the problems. Then, the main program of GNPsb-Sarsa can re-use the subroutine to enable a faster evolution and even better performances. GNPsb-Sarsa containing a main program and a subroutine evolves by natural selection and genetic operations, where the gene of GNPsb-Sarsa is the pair of the main GNP and its subroutine. That is, the genetic operations on GNPsb-Sarsa are constrained by the gene structure on which they can operate. The following two experiments are discussed: 1) Testing one subroutine node in the main GNP 2) Varying the number of subroutine nodes in the main GNP. The results of simulations showed that the proposed GNP sb-Sarsa can provide reasonable opportunities for evolving complex solutions.",
• abstract = "Chapter 3 introduces a methodology to enhance the generalisation ability of the stock trading models based on GNP-Sarsa with multi-subroutines (GNPmsb-Sarsa), which is a kind of extension of GNPsb-Sarsa. The proposed method is developed for discovering the nodes and node connections to realise functions, and the functional distributed modular GNP is developed. The important points of the subroutines mechanism are as follows: First, the nodes and node connections discovered in the subroutines are reused to create effective trading rule s for certain function. Second, the evolution can be achieved so quickly by narrowing the search space with subroutines. Last, as the kinds of functional subroutines increase, the generalization ability is improved since more generalised frequent transitions of GNP, i.e., building blocks are found instead of precisely modelling the training data, which leads to the over fitting problem. Simulation results showed that the proposed method can generate more efficient and generalised trading models and obtain much higher profits.

  Chapter 4 introduces a methodology to enhance the generalisation of the stock trading models based on Cooperative Coevolutionary Genetic Network Programming-Sarsa (GNPcc-Sarsa). The basic idea comes from both natural and artificial systems, which show that an integrated system consisting of several subsystems can reduce the total complexity of the system and solve a difficult problem satisfactorily. Therefore, a cooperative coevolution approach is proposed, where several species simultaneously evolve. Such an approach allows different species of the GNP-Sarsa model to evolve in a parallel and cooperative manner, which makes the generated model more robust, generalised and efficient for generating stock trading strategies. GNPcc-Sarsa places as few restrictions as possible to the structure, allowing the model to obtain a wide variety of architectures during the evolution and to be easily used to solve complicated problems. It has been found from simulations that the performances of the proposed model are better than those of other methods.

  Chapter 5 introduces a methodology to simultaneously learn several tasks based on GNP, which is called GNP with multitasks (GNPmt), where each GNP among several GNPs corresponding to several tasks is used to learn its own task. GNPmt has some features, such as distribution, interaction and autonomy, which are helpful for learning multitask problems. The experimental results on the self-sufficient collecting problem are given to illustrate that GNPmt can give BibTeX entry too long. Truncated

Genetic Programming entries for Yang Yang

Citations