摘要
When a high-dimensional and complicated system has an unknown structure, variables and parameters, the evolutionary algorithm is most often applied and achieves better results. In the analysis of these developed population-based evolutionary algorithms, current evolutionary algorithms do not have the ability to solve both parameter and combinatorial optimization problems. Three evolutionary algorithms with flexible crossover, composition abilities, and learning with memorizing functions are developed in this dissertation. System modeling and pattern recognition are used as the discussion targets for verifying the developed algorithms. The three novel evolutionary algorithms are (1)Accumulated Genetic Algorithm (A-GA): a simple variable structure genetic algorithm with accumulated learning to solve fuzzy PID controller design. The total number of parameters learned is accumulated process by process. (2)The modified Genetic Programming (GP): the GP can be treated as most direct variable structure evolutionary algorithms. There are two big flaws involving trees growth and parameter estimation in the GP. The GP proposed in this dissertation can overcome this difficulty using some novel operators to automatically find the minimal structure size. (3)Variable Structure Genetic Algorithm (VSGA): The conventional genetic algorithm (GA) is limited by the fixed chromosome size which prevents it from solving functional optimization problems like the GP. Since the idea behind the GA and GP is to simulate animal evolutionary behavior. The GA as a variable structure evaluation algorithm can be used to solve functional optimization problems. To improve GP and VSGA’s learning capabilities an effective directed initialization scheme, a pruning technique and reorganization approach, and a new operator called extinction and regeneration are proposed in this dissertation. Learning experience with memory is embedded into the VSGA to enhance its learning capability. Three application fields are used to verify the effectiveness and efficiency of the accumulated GA, the proposed modified GP and VSGA. (1)Fuzzy controller design: The gain scheduled fuzzy PID (GS_FPID) controller’s novel structure consists of a fuzzy PI-like controller and fuzzy PD-like controller. Both the fuzzy PI-like and PD-like controllers are weighted through adaptive gain scheduling, determined by fuzzy logic inference. The A-GA is designed to learn the fuzzy inference system parameters. In order to learn the number of fuzzy rules required for the TSK model, the fuzzy rules are learned in an accumulated way. In other words, the parameters learned in the previous rules are accumulated and updated along with the parameters in the current rule. It will be shown that the proposed GS_FPID controllers learned by the A-GA perform well for regular linear systems and also for higher order and time-delayed systems. (2)Signal process: A Volterra filter with high order and large memories contains a large number of cross product terms. Instead of applying evolutionary algorithms to search for all input signal cross products, it is utilized to search for a smaller set of primary signals that evolve into the whole set of cross products. The important primary signals and the associated cross product signals contributing most to the outputs are chosen while the primary signals and the associated input signal cross products that are trivial to the outputs are excluded from the possible candidate primary signals. (3)Pattern recognition: The proposed VSGA aims to select possible combinations with the smallest feature quantity while maximizing the classification accuracy. Besides the classification rate and number of features, evolution algorithm solution usability verification is discussed in this dissertation.
並列摘要
When a high-dimensional and complicated system has an unknown structure, variables and parameters, the evolutionary algorithm is most often applied and achieves better results. In the analysis of these developed population-based evolutionary algorithms, current evolutionary algorithms do not have the ability to solve both parameter and combinatorial optimization problems. Three evolutionary algorithms with flexible crossover, composition abilities, and learning with memorizing functions are developed in this dissertation. System modeling and pattern recognition are used as the discussion targets for verifying the developed algorithms. The three novel evolutionary algorithms are (1)Accumulated Genetic Algorithm (A-GA): a simple variable structure genetic algorithm with accumulated learning to solve fuzzy PID controller design. The total number of parameters learned is accumulated process by process. (2)The modified Genetic Programming (GP): the GP can be treated as most direct variable structure evolutionary algorithms. There are two big flaws involving trees growth and parameter estimation in the GP. The GP proposed in this dissertation can overcome this difficulty using some novel operators to automatically find the minimal structure size. (3)Variable Structure Genetic Algorithm (VSGA): The conventional genetic algorithm (GA) is limited by the fixed chromosome size which prevents it from solving functional optimization problems like the GP. Since the idea behind the GA and GP is to simulate animal evolutionary behavior. The GA as a variable structure evaluation algorithm can be used to solve functional optimization problems. To improve GP and VSGA’s learning capabilities an effective directed initialization scheme, a pruning technique and reorganization approach, and a new operator called extinction and regeneration are proposed in this dissertation. Learning experience with memory is embedded into the VSGA to enhance its learning capability. Three application fields are used to verify the effectiveness and efficiency of the accumulated GA, the proposed modified GP and VSGA. (1)Fuzzy controller design: The gain scheduled fuzzy PID (GS_FPID) controller’s novel structure consists of a fuzzy PI-like controller and fuzzy PD-like controller. Both the fuzzy PI-like and PD-like controllers are weighted through adaptive gain scheduling, determined by fuzzy logic inference. The A-GA is designed to learn the fuzzy inference system parameters. In order to learn the number of fuzzy rules required for the TSK model, the fuzzy rules are learned in an accumulated way. In other words, the parameters learned in the previous rules are accumulated and updated along with the parameters in the current rule. It will be shown that the proposed GS_FPID controllers learned by the A-GA perform well for regular linear systems and also for higher order and time-delayed systems. (2)Signal process: A Volterra filter with high order and large memories contains a large number of cross product terms. Instead of applying evolutionary algorithms to search for all input signal cross products, it is utilized to search for a smaller set of primary signals that evolve into the whole set of cross products. The important primary signals and the associated cross product signals contributing most to the outputs are chosen while the primary signals and the associated input signal cross products that are trivial to the outputs are excluded from the possible candidate primary signals. (3)Pattern recognition: The proposed VSGA aims to select possible combinations with the smallest feature quantity while maximizing the classification accuracy. Besides the classification rate and number of features, evolution algorithm solution usability verification is discussed in this dissertation.
參考文獻
延伸閱讀
- 洪偉展（2014）。運用次經驗演算法於分類程序之構建〔博士論文，國立暨南國際大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0020-0302201423394600
- Lin, Y. J. (2018). 改進針對基於模型之基因演算法的指數族群方法 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU201802888
- 夏偉中（2005）。多表現序列標籤到基因體排列演算法之設計與實作〔碩士論文，亞洲大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0118-0807200916271793
- 黃敬庭（2023）。Design of Evolutionary Algorithm for Solving Multimodal Continuous Optimization Problems〔碩士論文，國立臺灣師範大學〕。華藝線上圖書館。https://doi.org/10.6345/NTNU202300305
- Bhuvana, J. (2015). A Hybrid Evolutionary Algorithm for Discrete Optimization. Research Journal of Applied Sciences, Engineering and Technology, 9(9), 770-777. https://www.airitilibrary.com/Article/Detail?DocID=20407467-201503-201506090034-201506090034-770-777