Differential Evolution Algorithm And Application Research Of Multi-strategy Fusion

The evolutionary algorithm is one of the swarm intelligence optimization methods with global search performance,which is mainly used to solve non-convex,non-differentiable and discrete optimization problems.As a modified model of evolutionary algorithms,the differential evolution has a feature of heuristic search,because it is mainly based on a compound vector of directions between different points to generate new individual.It has some advantages,such as good genetic information fusion,simple structure and strong robustness,and can be adopted in solving common optimization problems.However,in the era of big data,the scale and structure of optimization problems from engineering and social practice are becoming larger and more complex than ever.Classic evolutionary algorithms,including differential evolution,can not solve these problems effectively in many complicated cases because they are trapped in local optimum.In order to overcome these shortcomings,for four common optimization models,such as global optimization,constrained optimization,multi-objective optimization and hierarchical optimization,based on the framework of differential evolution,this dissertation focus on improving algorithm performance by analyzing and using the problem-specific characteristics and evolutionary strategies.The main research results are as follows:1.In order to solve complex global optimization problems,a new differential evolution is designed by using the methods of multi-stage and multi-strategy.Firstly,in the early stage of evolution,the distribution information of decision variables in the decision space is adopted,and the oppositional strategy with disturbance is used for some worse individuals to enhance the exploration ability of the algorithm.Then,the elite sharing strategy is integrated in the generation of offspring,and high-quality individuals are expected to be produced through gene exchange with the better individuals.Furthermore,the multi-mutation hybrid strategy is applied in different evolutionary stages to balance the exploitation and exploration capabilities of the algorithm.Finally,a self-adaptive crossover probability and a mutation factor selection method are designed by using the significance test of individual fitness difference.When compared with similar representative algorithms,the proposed differential evolution shows better performance.2.For constrained optimization problems,there exists lots of compuataion difficiults when constraints are too tight.In order to overcome the shortcomings,based on multipopulation,multi-stage and constraint handling schemes,a new evaluation strategy is constructed in differential evolution.Firstly,in the early stage of evolution,some reference points are used to establish dynamic adjustment lines,and two distance indicators are used to divide the parent population into three subpopulations.In the late stage of evolution,the population is divided into two subpopulations by using feasibility conditions.Then,according to the individual characteristics in the subpopulation,the corresponding mutation operators are selected to carry out the mutation operation,respectively.Finally,according to the location of infeasible points around feasible region,these infeasible solutions are classified and further transformed into feasible ones,which can effectively improve the ratio of feasible solutions.The simulation results show that the multi-population co-evolution and the constraint handling technique are feasible and efficient.3.In order to solve efficiently multi-objective optimization problems,a multiobjective differential evolution is proposed by using locally optimal individual information and adopting an external archiving scheme.Firstly,locally optimal individuals are used in the mutation operation,and a heuristic crossover operator is established by using the uniform design to produce high-quality crossover offspring.Then,weight values are taken to set up an external archive which stores some potential individuals,and these stored individual are randomly selected to construct an alternative mutation operation.In addition,the adaptive adjustment strategy of variation factors is established by using objective values to balance the convergence and diversity.Compared with state-of-the-art multi-objective optimization approaches,the proposed algorithm provides more high-quality Pareto optimal solutions than other compared approaches.4.For the deployment problem of virtual network function chain of elastic optical network among data centers,a new bilevel optimization model is established,and a hybrid approach is developed by combining a distributed estimation algorithm and a differential evolution.Firstly,according to the characteristics of the upper-level variables,the upper-level objectives are optimized by using a binary-encoded distribution estimation algorithm.Then,the differential evolution is used to solve the lower-level problem,and a correlation coefficient scheme is utlized to select individuals in the design of mutation operator,and in the operation some high-quality individuals are embedded to produce mutation offspring.In this way,the search efficiency of the algorithm can be evidently improved.The experimental results on two real-world examples show that the proposed algorithm provides a better deployment scheme.