Research On Evolutionary Large-Scale Optimization Algorithms

In the era of data and intelligence,there are a large number of optimization problems abounding in the production and living practices of the society.Such problems often have complex properties,such as non-convex,multimodal,and non-differentiable,and even difficult to establish an exact mathematical model.Evolutionary algorithms are widely used for solving complex optimization problems because they employ population search.However,the proliferation of problem dimensions poses a serious challenge to the scalability of evolutionary algorithms.The research topic of this dissertation is evolutionary large-scale optimization algorithms,and focuses on the cooperative coevolution framework for unconstrained and constrained large-scale optimization problems,the fundamental strategies of evolutionary optimization,and the design of large-scale optimization benchmarks.Specifically,the main research contents and innovations of this dissertation are summarized as follows.(1)A difficulty and contribution-based cooperative coevolution framework is proposed for large-scale unconstrained optimization problems,which allocates more computational resources to components with larger contribution and greater difficulty.Moreover,a dynamic difficulty estimation method is designed to quantify the problem difficulty based on the fitness landscape as well as the algorithm behaviors.Experimental results show that the proposed framework outperforms the contribution-aware cooperative coevolution framework and outperforms the outstanding algorithms in the CEC large-scale optimization competitions,such as the winner of the CEC’2010 competition,MA-SW-Chains,and the winner of the CEC’2013 and CEC’2015 competitions,MOS-CEC2013.(2)A constraint-objective cooperative coevolution framework(COCC)is proposed for large-scale constrained optimization problems,which allocates computational resources based on the influence of components on objective values and constraint violation.Secondly,a set of small-scale constrained optimization problems is extended into a large-scale continuous constrained optimization benchmark based on the existing guidelines for designing general large-scale optimization benchmarks.Finally,three evolutionary algorithms are selected as optimizers for the COCC framework and tested on the proposed benchmark.The experimental results demonstrate that COCC significantly outperforms several evolutionary large-scale constrained optimization algorithms.(3)The fundamental strategies of evolutionary optimization are studied.First,a density-based population initialization strategy is proposed to generate an initial population with both uniformity and randomness to enhance the performance of the evolutionary algorithm.Second,a random direction repair is proposed to guide the infeasible solution to move along the feasible domain to search the feasible solution.Third,a hybrid clustering strategy is proposed to enhance the clustering efficiency and exploration efficiency of the brain storm optimization.Fourth,a hybrid of particle swarm optimization and evolutionary strategy is proposed,which integrates the advantages of particle swarm optimization to explore the global landscape and the ability of evolutionary strategy to exploit local landscapes.The experimental results show that the fundamental strategies proposed in this paper could enhance the performance of evolutionary algorithms.(4)A new design guide for large-scale optimization benchmarks is proposed,which incorporates two new features abstracted from real-world large-scale design optimization problems:heterogeneous design and versatile coupling,based on the existing largescale optimization benchmarks.Then,a new large-scale optimization benchmark suite is designed including 15 large-scale optimization problems.Four representative algorithms are selected and tested on the proposed benchmark,and the results show that the existing evolutionary large-scale optimization algorithms are difficult to handle largescale optimization test benchmarks with heterogeneous and versatile coupling modules.Overall,the research on evolutionary large-scale optimization algorithms for largescale optimization problems in this dissertation has achieved innovative results.