Multi-method Ensemble Evolutionary Algorithms

As the performance rising, Evolutionary algorithm (EA) has become more andmore important in optimization domain. Compared with classical optimization meth-ods, EAs require less parameter settings and less pre-knowledge. For the engineeringapplications with many constrains, EAs can work without complicated uniformizationprocedure. Besides, EAshaveshownespeciallybetterperformanceonmulti-modalandlinked problems.Although EAs have achieved significant performance previously, several aspectsof EAs still need to be improved. Generally speaking, there are three demerits of EAsas follows: (1) their universality, which means the ability of solving diverse kinds prob-lems, is still relatively poor; (2) their scalability, which is mainly the capability of solv-ing high dimensional problems, needs to be improved; (3) their application in engineer-ing optimization problems is still limited. For the first two drawbacks, we design amore robust and effective EA framework by using multi-method ensemble idea. Fur-thermore, the proposed EAs are applied to solve real world engineering applicationsand scientific problems, and the performance mostly surpass those of the classical EAs.This dissertation focuses on designing more effective and efficient multi-methodensemble based EAs, and then applying them to difficult engineering and scientificoptimization problems. The important findings are as follows:1. InordertoimprovethescalabilityofEstimationofDistributionAlgorithm(EDA),we design a new self-adaptive mixed distribution based sampling operator andthen,proposeaself-adaptiveMixeddistributionbasedUni-variateEDA(MUEDA).The advantages of MUEDA compared with the classical EDAs and state-of-the-art EAs are verified on the function optimization experiment scaling from 30 to1500 dimension.2. In order to improve the universality of Particle Swarm Optimization (PSO), wepropose a self-adaptive learning framework to extract strengthes from differentPSO new offspring creation strategies, which results in a new PSO variant self-adaptive learning PSO (SLPSO). Generally, SLPSO can assign more computa-tionalbudgettosuitablestrategiesbasedonthefeedbackofpreviousoptimizationprocedure. In this case, the universality of SLPSO can be remarkably improvedcompared with the other PSOs. This viewpoint is confirmed by function opti- mization and economic load dispatch of power system experimental results.3. In the previous research, the parallel execution of multiple new offspring cre-ation strategies is the usual form of multi-method ensemble algorithms. In thiskind of algorithmic framework, the learning mechanism is crucial to extend theapplication area. Therefore, it is difficult to handle ill-conditional and deceiv-able problems. In order to solve this, we propose a new Two-Stage based serialmulti-method ensemble EA (TSEA) framework, whose main idea is to dividethe optimization procedure into two stages, global convergence and exhaustivesearch.Based on TSEA framework, we propose a series of algorithmic versions to han-dle general function optimization problems, large scale optimization problems, multi-objective optimization problems and dynamic multi-objective optimization problems.Besides the function optimization experiments, TSEA shows significant advantages onreal world engineering and scientific optimization applications:1. Large scale global optimization: (with more than 102 variables) MUEDA andTSEA provide much better performance than the current state-of-the-art largescaleglobaloptimizationalgorithms. Asthedifficultylevelarises,theadvantagesof our proposed algorithms in effectiveness and efficiency become more clear. InIEEE CEC2008and2010largescaleglobal optimization competitions, MUEDAand TSEA are among the best candidates.2. Large scale ELD optimization: As an important and difficult optimization taskin power system, ELD has attracted public attention from research community.However,theperformanceofpreviousalgorithmsisstilllimited. Inordertosolvethis, weapplySLPSOandaTSEAversion, ED-DE,tolargescaleELDproblems.ComparedwiththecurrentbestELDoptimizationalgorithms,ED-DEcanallbestsolution records within lower computational cost.3. Digital IIR filter design: In the previous research, many methods have been ap-plied to this optimization task. However, they have similar demerits: (1) theycan only be applied to low-quality filter design; (2) the filter is represented byfloating-point numbers. These two aspects require more effective design meth-ods. This dissertation adopts Two-Stage based multi-method ensemble memeticAlgorithm (TSMA) to solve this problem and achieve remarkable progresses. It is worth noting that TSMA shows especially reliable performance on the hardtasks, on which the other algorithms all fail.Besides, the TSEA based algorithm is applied to dynamic multi-objective optimizationdomain, which attracts little attention previously, and gain significant progresses. Insummary, TSEA framework shows promising performance in terms of effectiveness,efficiency, robustness and universality, and is especially suitable for hard optimizationtasks.