Large-Scale Multi-Objective Evolutionary Optimization Algorithms And Applications

Multi-objective Optimization Problems(MOPs)are a type of challengeable op-timization problem that exists widely in scientific research and production applica-tions.The difficulty lies in the conflict between multiple objectives.As there is usu-ally no solution that optimizes all the objectives,existing analytical methods cannot effectively solve these problems.Evolutionary Algorithms(EAs)naturally conform to the need for multi-objective optimization to find a set of approximations of an opti-mal set of solutions that compromise between multiple objectives.This is due to their population-based search mechanism,and that they do not require special assumptions about the problem characteristics.Therefore,Multi-Objective Evolutionary Algorithms(MOEAs)have become one of the main means of solving such problems.With the development of MOEAs,their scalability have gradually attracted the attention of re-searchers.Compared with the extensive attention of the multi-objective evolutionary algorithm in the scalability of the number of objective functions,scalability in deci-sion variable space is a topic that has been only scarcely studied in the context of MOEAs.However,this does not meet the actual needs.First,many real-world ap-plications are actually large-scale MOPs,such as the weight optimization in deep neu-ral network,large-scale social network multi-objective analysis,multi-objective vehi-cle routing problems,etc.Second,the performance of the existing MOEA decreases sharply with the increase of the number of decision variables,which makes it difficult to effectively solve large-scale MOPs.Motivated by this,the goal of this dissertation is to study large-scale multi-objective evolutionary optimization.To achieve this goal,this dissertation analyzes the characteristics of continuous and discrete MOPs and designs effective MOEAs for them.Specifically,the main research contents and contributions of this dissertation can be summarized into the following three aspects:1.An analysis of the challenges associated with large-scale multi-objective opti-mization is carried out and a new large-scale MOEA is proposed based on the analysis.When solving MOPs,one usually seeks a set of Pareto-optimal sets to achieve a compromise between multiple objectives.MOEAs have been widely proposed for this purpose.However,their performance often deteriorates rapidly as the number of decision variables increases.Therefore,the exclusive challenges along with the increase of the number of variables of a MOP are examined empir-ically,and the popular benchmarks are categorized into three groups accordingly.The three groups are convergence-focused,diversity-type I and diversity-type II.Intuitively,convergence-focused problems can be mitigated using techniques em-ployed in large-scale single-objective optimization.Diversity-type I problems require MOEAs to have stronger diversification but ignore a correlation between position and distance functions.Available evidence indicates that this type of problems can be solved well using existing large-scale MOEAs.The rest of the problems pose a great challenge to the balance between diversification and con-vergence by considering a correlation between position and distance functions.While existing large-scale MOEAs perform well on the problems in the first two categories,they suffer a significant loss when applied to diversity-type ? prob-lems,and even perform worse than classical MOEAs.This spurs our further investigation,and a large-scale MOEA with an enhanced diversification mecha-nism is proposed.It employs a new solution generator with an external archive to improve the diversification ability,using SMS-EMOA as the basic framework.The main idea of the new solution generator is to encourage diverse solutions by forcing the search towards different parts of the Pareto front.A dual local search mechanism is proposed to implement this.Experimental results demonstrate that it outperforms compared algorithms on diversity-type II problems and show its advantage in the balance between diversification and convergence.2.A complex real-world large-scale multi-objective continuous optimization prob-lem,namely ROC convex hull maximization problem,is studied.The funda-mental tradeoffs between the conflicting TPR and FPR permit this problem to be cast as a MOP.As the number of variables,such as features and classifier hyper-parameters,increases dramatically,such problem exhibits the characteristics of large-scale MOPs.Nevertheless,it is different from general MOPs,as its goal is to seek a set of optimal solutions called convex hull solutions.For this special MOP,researchers have proposed some convex hull-based MOEAs.However,available evidence indicates that they exhibit a diversity loss when solving large-scale ROC convex hull maximization.To deal with this issue,an improved convex hull-based MOEA is proposed to improve the diversification ability.Specially,convex hull-based sorting with convex hull of individual minima and extreme area extraction selection are proposed as a novel selection operator.Empirical studies on multi-ple high-dimensional and imbalanced datasets show that it outperforms compared algorithms for evolving neural networks.3.A complex real-world large-scale multi-objective discrete optimization problem,namely influence maximization in large-scale social network,is studied.It is a combinatorial optimization problem that has received widespread attention in re-cent years.In a competitive environment,the study of this problem aims to single out a set of seeds with the least cost and its influence exceeds a specified thresh-old under the competition of other participants.However,as the scale of social networks increases,existing methods could hardly provide a quality acceptable solution in an acceptable time.Motivated by this,this paper firstly models the influence maximization problem in the competitive environment as a MOP,and then designs a scalable MOEA to solve it.Specifically,a variable feasible-region search is proposed to enhance the search ability to improve the balance between diversity and convergence.Experimental results on multiple large-scale social networks show that the proposed algorithm can provide a better balance between the influence performance and the time overhead.