Note (General)type:Thesis
In experiments and applications, optimization problems aim at finding the best solution from all possible solutions. According to the number of objective functions, optimization problems can be divided into single-objective problems and multi-objective problems. In this thesis, we focus on solutions for single-objective optimization problems. The purpose of this thesis is to clarify a means for realizing high search accuracy without parameter adjustment.To achieve high accuracy results for single-objective optimization problems, there are four major points to note: the local search ability in unimodal problems, the global search ability in multimodal problems, diverse search patterns for different problems, and the convergence speed controlling. Population-based methods like the particle swarm optimization (PSO) algorithms are often used to solve single-objective optimization problems. However, the PSO is a parameter-needed method which means it needs to adjust parameters for better performances. The adjustment of parameters becomes an overhead when considering for engineering applications. Besides, the bare bones particle swarm optimization (BBPSO) algorithm is a parameter-free method but unable to change the search pattern according to different problems. Also, the convergence speed of the BBPSO is too fast to achieve high accuracy results. To cross the shortcoming of existing methods and present high accuracy results for single-objective optimization problems, seven different hierarchical strategies are combined with the BBPSO in this thesis. Four of the proposed algorithms are designed with swarm division which are able to converge to the global optimum fast. The other three algorithms are designed with swarm reconstruction which are able to slow down the convergence and solve shifted or rotated problems. Moreover, no parameter adjustment is needed when controlling the convergence speed.First of all, four algorithms with swarm division are proposed. In the pair-wise bare bones particle swarm optimization (PBBPSO) algorithm, the swarm splits into several search units. Two particle are placed in one unit to enhance the local search ability of the particle swarm.To increase the global search ability, the dynamic allocation bare bones particle swarm optimization (DABBPSO) algorithm is proposed. Particles in DABBPSO are divided into two groups before evaluation according to their personal best position. One group is named as the core group (CG) and the other one is called the edge group (EG). The CG focuses on digging and trying to find the optimal point in the current local optimum. Conversely, the EG aims at exploring the research area and giving the whole swarm more chances to escape from the local optimum. The two groups work together to find the global optimum in the search area.To solve the shifted of rotated problems, traditional methods usually need to increase the population size. However, the growth of population size may increase the computing time. To cross this shortcoming, a multilayer structure is used in the triple bare bones particle swarm optimization (TBBPSO) algorithm. The TBBPSO is able to present high accuracy results in shifted and rotated problems without the increasing of population size.
In real-world applications, optimization methods are required to solve different types of optimization problems. However, the original BBPSO can not change its search pattern according to different problems. To solve this problem, a bare bones particle swarm optimization algorithm with dynamic local search (DLS-BBPSO) is proposed. The dynamic local search strategy is able to provide different search patterns based on different questions.In engineering applications, the optimization results can be improved by controlling the convergence speed. Normally, traditional methods need parameter adjustment to control the convergence speed. It is difficult to adjust the parameters for every single problem. To solve this problem, three different reorganization strategies are combined with the BBPSO. In the bare bones particle swarm optimization algorithm with co-evaluation (BBPSO-C), a shadow swarm is used to increase the diversity of the original swarm. A dynamic grouping method is used to disperse both the shadow particle swarm and the original particle swarm. After the dispersion, an exchanging process will be held between the two swarms. The original swarm will be more concentrated and the shadow swarm will be more scattered. With the moving of particles between the two swarms, the BBPSO-C gains the ability to slow down the convergence without parameter adjustment.With the improvement of technologies, it is possible to get high accuracy results with a long calculation. In the dynamic reconstruction bare bones particle swarm optimization (DRBBPSO) algorithm, a dynamic elite selection strategy is used to improve the diversity of the swarm. After elite selection, the swarm will be reconstructed by elite particles. According to experimental results, the DRBBPSO is able to provide high accuracy results after a long calculation.To adapt to different types of optimization problems, a fission-fusion hybrid bare bones bare bones particle swarm optimization (FHBBPSO) is proposed. The FHBBPSO combines a fission strategy and a fusion strategy to sample new positions of the particles. The fission strategy aims at splitting the search space. Particles are assigned to different local groups to sample the corresponding regions. On the other side, the fusion strategy aims at narrowing the search space. Marginal groups will be gradually merged by the central groups until only one group is left. The two strategies work together for the theoretically best solution. The FHBBPSO shows perfect results on experiments with multiple optimization functions.To conclude, the proposed hierarchical strategies provide each of the BBPSO-based algorithms variants with different search characteristics, which makes them able to realize high search accuracy without parameter adjustment.
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2020-07-06T20:31:19+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション