The Multi-objective Ensemble Optimization Algorithm And Its Application

Multi-objective optimization algorithm assists human being to make the optimal decision making under the two or more conflicting objectives.It is an important research direction of intelligent computing.With the increasing number of researchers,a series of problems in the theory and system of multi-objetcive optimization are well solved,such as many objective optimization problem,large-scale decision variable problem and dominat level problem.The success of multi-objective optimization in theory contributes researchers to focus much attention on solving the applied engineering problems.But,considering that the large gap between engineering model and benchmark function,researchers design the specific optimization framework,optimization objectives,searching algorithms,the termination conditions and deceision-making scheme for efficiently solving the multi-objective optimization problem in applied engineering.Our paper concludes this multi-objective engineering-oriented problem as multi-objective efficient optimization problem.Our paper demonstrates the difficulty and challenges in multi-objective efficient optimization problem on two aspects,which includes multi-objective machine learning model’s efficient optimization problem and multi-objective antenna array distribution’s efficient optimization problem.On the first aspect,general machine learning model normally implements task under two conflicting objectives,namely performance and cost.For example,feature selection needs to use the smallest number of features to obtain the highest accuracy.Benefiting from population-based searching and multi-objective dominant relationship,multiobjective algorithms efficiently seek the optimal parameters under two conflicting objectives,which assists machine learning model to be a compromising situation between performance and cost.On the second aspect,antenna array distribution problem not only balances performance and cost,but also designs the efficient deceision making scheme to give the optimal antenna array distribution to engineer.Benefiting from flexible optimization framework and fuzzy experience deceision making,multi-objectives optimization algorithms rapidly seeks the pareto antenna array distribution according to the actual situation,then recommend the optimal feasible scheme to engineers.According to these two aspects,our paper introduces five challenges for illustrating the multi-objective efficient optimization solutions,which includes(1)In multi-objective fuzzy relationship classifier,the imbalanced problem between classification process and clustering process is urgent to be solved.(2)In multi-objective large-scale feature selection algorithm,the critical problem is the compromising solution simultaneously for feature number’s minimizing and classification accuracy’s maximizing.(3)In multi-objective dynamic domain adaption,the dynamical weight trajectory between marginal distribution difference and conditional distribution difference is urgent to be solved.(4)In the multi-objective preselection-based algorithms,the joint consideration of evaluation number and solutions quality is the important improblem to be solved.(5)In multi-objective sparse span antenna,the critical problem is the imbanlaced problem between antenna’s number and array’s gain.The main contributions of this paper are highlighted in the following:(1)Aiming at image classification’s multi-objective efficient optimization problem,which is the seeking optimal parameters in fuzzy relationship classifier,our paper proposes a dynamic simultaneous clustering and classification algorithms via automatic differential evolution and firework operators.Firstly,a novel framework of multi-objective optimization integrating with fuzzy relationship classifier(SCC-MOEA)is proposed.Then,a combination searching strategy for dynamic searching the optimal parameter is proposed.Meanwhile,a rapid and lowcomplexity silhouette coefficient is redesigned as clustering objective function.Finally,automatic clustering,opposition-based learning and adjusted mutual information are presented for strengthening SCC-MOEA framework.(2)Aiming at feature selection’s multi-objective efficient optimization problem,which is the seeking optimal feature subsets in feature selection tasks,our paper proposes two multiobjective large-scale feature selection algorithms.The first scheme is a dividing-based manyobjective evolutionary algorithm for large-scale feature selection(DMEA-FS).Firstly,four novel objectives are established for exploring the optimal feature’s subsets.Meanwhile,we design two structures of wrapper for high accuracy and filter for low computation cost in DMEA-FS.Secondly,two new recombination methods are presented for rapid convergence.Mapping-based variable dividing is presented for precise related variables.Thirdly,based on minimum Manhattan distance,a triangle-approximating decision-making is proposed for assisting users’ determination with/without preference information.The second scheme is a novel multi-objective large-scale cooperative coevolutionary algorithm for three-objectives feature selection is proposed,termed MLFS-CCDE.Specially,three novel objectives are established for exploring the optimal feature’s subsets.Meanwhile,the wrapper structure of MLFS-CCDE is designed for large-scale feature selection and effective diversity searching.Moreover,cluster-based decomposition strategy is elaborated for reducing the computation in decision variable’s decomposition and dual indicator-based representatives are elaborated for balancing the convergence and diversity of representative solution.Finally,a high-efficiency and accurate heart disease diagnosis system based on MLFS-CCDE framework is constructed in cardiology.(3)Aiming at transfer learning’s multi-objective efficient optimization problem,which is the seeking optimal feature space in data transferring tasks,our paper proposes a novel multiobjective dynamical distribution adaption(MODDA)with instance reweighting.Firstly,a novel framework of multi-objective optimization integrating with domain adaption is proposed,which narrows the discrepancies of marginal distribution and conditional distribution.Specially,a customized NSGA2 optimization method is presented for searching the optimal path of cumulative weight.Then,four combinations of genetic operators are compared for constituting MODDA.Finally,kernel mean matching(KMM)is firstly presented for dynamic compensation depending on individual’s relevance in instance reweighting.(4)Aiming at algorithm-oriented multi-objective efficient optimization problem,which is the seeking optimal parameters in MOEA’s preselection process,our paper proposes an indicator-based evolutionary algorithm based on pre-selection with classification guided by SVM.Firstly,a novel preselection approach to improve the performance of the IBEA is presented,in which a SVM(Support Vector Machine)classifier is adopted to sort the promising solutions from unpromising solutions and then the newly generated solutions are conversely added as train sample to increase the accuracy of the classifier.Finally,an online and asynchronous training method for SVM model with empirical kernel is presented.(5)Aiming at resource allocation’s multi-objective efficient optimization problem,which is the seeking optimal array distribution in antenna array design,our paper proposes two multiobjective sparse span antenna distribution algorithms.The first scheme designs a multiobjective self-organizing optimization for constrained sparse array synthesis,termed MOSSA.Overall,a uniform framework of multi-objective sparse span array is proposed.Specially,two objectives,number of selected antenna and peak side lobe level,are established for exploring the optimal array distribution in the framework.Based on the framework,for the problem of global-optimum array distribution,we propose a multi-objective particle swarm optimization searching pattern and design a MOSSA algorithm;Furthermore,for the problem of flexiblyadjusted self-organizing array structure,we present a multiobjective genetic programming searching pattern and design a MOSSA-gp algorithm.Moreover,a limited-region mode supplements to the framework.Finally,combination decision strategy assists users to screen out suitable solutions under the guidance of fuzzy-range indexes and then select the optimal solution by a triangleapproximating approach based on minimum Manhattan distance.The second scheme designs a three-objectives sparse span antenna distribution algorithm.Firstly,antenna’s number,array gain and null beam width is designed as optimization objective.Then,a strengthed NSGA3 is presented for searching the optimal antenna distribution.Finally,a brushing technique with multiple-information-point to assist the decision maker for selecting the high-performance solution.