Font Size: a A A

Multiobjective Optimization Evolutionary Algorithms Based On Local Learning And Uniform Decomposition

Posted on:2015-06-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:X L MaFull Text:PDF
GTID:1108330464968902Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
The optimization problem is a basic issue in current engineering application and scientific research. If the optimization problem has only one objective function, it is called single-objective optimization problem. When the number of objective functions is larger than one, it is called multi-objective optimization problems (MOPs). To deal with MOPs, there are two kinds of methods:decomposition technique of mathematical programming and evolutionary algorithm. Multiobjective optimizaition evolutionary algorithm based on decomposition (MOEA/D) combines both methods to solve for MOPs. It firstly decomposes an MOP into a set of single objective optimization subproblem and then uses evolutionary algorithm to evolve these subproblems simultaneously. Recently, MOEA/D has been widely studied owing to its good performances. This dissertation aims to improve the performance of MOEA/D with local learning and uniform decomposition techniques from the following aspects, including Baldwinian learning based on regularity property of MOP, opposition-based learning strategy in convergence accelerating, many-objective optimization problems, analyzing the features of decision variables. The ma^n contributions of the thesis can be summarized as follows:1. Machine learning technique has been integrated into multiobjective evolutionary algorithm. The traditional reproduction operators, which are originally designed for single-objective optimization, are directly adopted in most state-of-the-art multi-objective evolutionary algorithms (MOEAs). However, classical reproduction operators designed for SOPs might not be suitable for MOPs due to the different optima structures between them. Regularity property of MOP that the PS of a continuous MOP is a piecewise continuous (m-1)-dimension manifold is introduced to deal with the issue and MOEA/D with Baldwinian learning based on regularity property of MOP (MOEA/D-BL) is proposed. MOEA/D-BL firstly uses local principal component analysis (PCA) to obtain the evolving information based on the learned distribution model of current population. It constructs a candidate descent direction based on the learned distribution model and the evolving history of the parent individuals. Comparative experiments show that the proposed Baldwinian learning operator can accelerate the convergence of solutions.2. To improve the performance of MOEA/D, we use opposition-based learning (OBL) to redesign population initialization and evolutionary operator and propose novel MOEA/D based on opposition-based learning (MOEA/D-OBL). The main idea behind OBL is to consider an estimate and its corresponding opposite estimate synchronously in order to obtain a better approximation for the optimal solution. A new population initialization based on OBL and opposition-based learning strategy are proposed in MOEA/D-OBL. Aiming to accelerate the convergence speed of the parent algorithm MOEA/D, opposition-based learning strategy combines the evolution operator and opposition-based local search. A lot of experimental studies have demonstrated that the effectiveness of OBL.3. Aiming to deal with many-objective problems, a modified Tchebycheff decomposition approach and novel weight vectors based on uniform decomposition measurement are integrated into MOEA/D. Thereby MOEA/D with uniform decomposition measurement (MOEA/D-UDM) has proposed. The modified Tchebycheff decomposition approach is used to find evenly scattered solutions over Pareto optimal front (PF). The novel weight vectors based on uniform measurement are constructed to obtain any amount of uniformly distributed weight vectors. A lot of experimental studies on many-objective problems have showed the effectiveness of the proposed algorithm.4. By learning the interactions among decision variables, each objective function of multiobjective optimization problem is decomposed into a set of simpler sub-functions with low-dimensional subcomponent. Therefore, the difficulty of optimization problem is reduced. By analyzing the control property of decision variable, the conflict of MOP can be extracted and the efficiency of algorithm can be improved. A multiobjective evolutionary algorithm based on decision variable analyses (MOEA/DVA) is proposed. A lot of experimental studies have demonstrated that the effectiveness of MOEA/DVA especially for complex and difficult MOPs.5. Aiming to introduce the preference information into the framework of MOEA/D, MOEA/D with biased weight adjustment inspired by user-preference (pMOEA/D) is proposed. pMOEA/D is used to solve the multi-objective reservoir flood control problem. In order to obtain uniformly distributed solutions over PF, pMOEA/D uses modified Tchebycheff decomposition instead of Tchebycheff decomposition. To focus the search for the interesting region of decision maker, some subproblems, which are far away from the preference regions, are deleted. And then some new subproblems, which are expected to search the preference regions, are added into the current evolutionary population. A lot of experimental studies have demonstrated that the effectiveness of pMOEA/D especially for multi-objective reservoir flood control problem.
Keywords/Search Tags:Multi-objective Optimization, Decomposition Strategy, Uniform Decomposition Measurement, Decision variable analyses, Regularity Property, Baldwinian Learning, Opposition-Based Learning
PDF Full Text Request
Related items