Font Size: a A A

Research On The Method Of Many-objective Optimization And The Evaluation About Redundancy Of The Objectives

Posted on:2015-08-12Degree:DoctorType:Dissertation
Country:ChinaCandidate:J ZouFull Text:PDF
GTID:1108330464471604Subject:Applied Mathematics
Abstract/Summary:PDF Full Text Request
Evolutionary algorithm is to simulate the process of biological evolution and search the global optimal solutions, which has been applied into the field of multi-objective optimization, and Multi-objective evolutionary optimization has been gradually developed into a frontier hot issue in that field. Recently, there are many up-to-date features of this issue proposed. For example, due to the complexity of the real-world optimization problems, it is difficult for the decision makers to distinguish which objective or indicator is important or not. Thus, all of them are considered. Farina and Amato gave a definition of many-objective optimization problems whose number of objectives is no less than four. In the real world, most of the optimization problems are multi-objective optimization problems. As we all known, the Time Table problem, belongs to the NP-hard problems, is also a class of many-objective optimization problems.A multitude of scholars domestic and abroad, have been researching on those problems but these problems are still not solved theoretically and technically.Recently, the research of high-dimensional multi-objective evolutionary algorithms is still not mature in domestic and abroad, the main research of which focus on the improvement of existing EMOs. Whether relaxation Pareto dominance relations, the Pareto dominated sorting and the way of objectives reduction are strong linked to some difficulties like parameters, bad convergence and diversity, high time complexity and reservation of the boundary individuals and so on. Nevertheless, on the theory and technology, there is no general method to solve the many-objective optimization problems.This paper, by fully analyzing different many-objective optimization algorithms in domestic and acknowledging the advantages and disadvantages between these algorithms, proposed two kinds of many-objective optimization algorithms and a metric about redundancy of the objectives as well as applied these methods into the eigenvalue extraction.This paper proposes a clustering technology in hyper-plane used to solve many-objective optimization problems without parameters, which is called ClusterISEA. By means of clustering technology, the algorithm categorizes the individuals on the critical layer. In the process of optimization, it does not introduce any parameters and make itself having strong ability to adapt to different problems. By experimenting on DTLZ series with another seven relevant algorithms, the proposed algorithm is proved to be effective.This paper proposes a rotation grid technology used to solve the many-objective optimization problems, which is called RGridEA. The algorithm uses the way of rotating grid to divide the objective space, which can obtain good diversity by the strategy of grid, and achieve good convergence and distribution by means of strategy of rotation mechanism which can separate the convergence and distribution, as well as introducing strategy of directions guidance. By experimenting on DTLZ series with another seven relevant algorithms, the proposed algorithm is proved to be effective.This paper, by approach of the idea of fitting to distinct the similarities of the distribution between the redundant objectives and non-redundant objectives in objective space, thus proposed a standard (OSDSR) to measure the pros and cons of the redundant objectives. This method takes the information between the redundant objectives set and non-redundant objectives set into account. Moreover, it applies subsection fitting mechanism to calculate the similarity degree of different objectives. By means of the calculation of similarities between the objectives, OSDSR has better accuracy. In terms of the experimental results on Dtlz5, Dtlz2BZ and Dtlz2, it is obvious that OSDSR can obtain good performance as well as IGD without any knowledge of Pareto front.In the end, this paper applies the proposed algorithms into the extraction of text eigenvalue, and the experimental results show that the algorithms can effectively reduce the dimensions of the eigenvalue and improve the performance of text categorization.
Keywords/Search Tags:Multi-objective evolutionary, Many-objective evolutionary, Multi-objective optimization, Optimization, Feature extraction, Text processing, Evolutionary algorithm
PDF Full Text Request
Related items