WO2022007376A1 - 基于贝叶斯自适应共振的多目标多模态粒子群优化方法 - Google Patents

基于贝叶斯自适应共振的多目标多模态粒子群优化方法 Download PDF

Info

Publication number
WO2022007376A1
WO2022007376A1 PCT/CN2021/070103 CN2021070103W WO2022007376A1 WO 2022007376 A1 WO2022007376 A1 WO 2022007376A1 CN 2021070103 W CN2021070103 W CN 2021070103W WO 2022007376 A1 WO2022007376 A1 WO 2022007376A1
Authority
WO
WIPO (PCT)
Prior art keywords
particle
population
particles
dominated
objective
Prior art date
Application number
PCT/CN2021/070103
Other languages
English (en)
French (fr)
Inventor
杨顺昆
姚琪
Original Assignee
北京航空航天大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京航空航天大学 filed Critical 北京航空航天大学
Publication of WO2022007376A1 publication Critical patent/WO2022007376A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/10Geometric CAD
    • G06F30/15Vehicle, aircraft or watercraft design
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/25Design optimisation, verification or simulation using particle-based methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/004Artificial life, i.e. computing arrangements simulating life
    • G06N3/006Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]

Definitions

  • the invention relates to the technical field of optimization algorithms, in particular to a multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance.
  • Some cluster-based particle swarm optimization algorithms often need to pre-set the number of clusters according to the number of solutions, but in reality it is difficult to know in advance how many solutions there are for multi-objective multi-modal functions, so the past methods generally take different Experiments are carried out with the number of clusters, and the optimal number of clusters is selected according to the experimental results, which brings great uncertainty and problem dependence to the optimization problem.
  • the present invention provides a multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance, the purpose of which is to obtain as many target spaces as possible in order to better solve the multi-objective multi-modal optimization problem
  • the Pareto frontier and the Pareto optimal solution set of the decision space ensure the diversity of solutions in the target space and decision space.
  • the multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance includes the following steps:
  • Sort the particles of each population according to the non-dominated sorting method and the special crowding distance according to the non-dominated sorting method, hierarchically sort all the particles in the various groups in the target space, and sort all the particles in the first non-dominated layer after sorting.
  • the particles of the layer are sorted in descending order according to the size of the special crowding distance;
  • the particles of each population are evolved based on the global optimal particle swarm optimization algorithm
  • the decision space of the particle is the space composed of the value range of each dimension of the particle; the target space is the space composed of the value range of the objective function determined by the decision space.
  • Non-dominated in S2 means: in a multi-objective problem, if one solution A is better than another solution B in all objectives, then solution A dominates solution B; if one solution A is not dominated by other solutions, Then the solution A is called a non-dominated solution.
  • the crowding distance is an index of the crowding degree between the target solution of a particle and the target solutions of its adjacent particles in the target space.
  • the specific content of S1 includes:
  • the calculation method of the posterior probability of the particle for the existing population is:
  • K represents the number of population existing
  • y l j represents the populations
  • the estimated prior probability of the jth population represents the possibility of population j to particle x;
  • ⁇ j represent the estimated mean vector and covariance matrix of the j population
  • the population sample size is represented by the hypervolume S J , which is the determinant of the Gaussian covariance matrix.
  • the hypervolume is the product of variances in each dimension:
  • d represents the particle dimension, represents variance
  • M J represents the number of particles in the J population
  • I represents the identity matrix
  • the non-dominated sorting method includes the following:
  • the target solution of these particles is divided into the second non-dominated layer in the target space;
  • step (2) all particles in the population are sorted by layers.
  • the calculation method of the special crowding distance is as follows:
  • M represents the number of all objective functions
  • CD im, obj represent the crowding distance of particle i in the dimension of objective function f m (x)
  • the calculation method is:
  • f m (x i+1 ) and f m (x i-1 ) represent the adjacent target solutions of the objective function solution f m ( xi ) of particle i, and f m (x max ) and f m (x min ) Represents the maximum and minimum values of the objective function f m (x);
  • CD i, obj ′ represents the crowded distance of particle i in the target space after normalization
  • CD i, obj ′ represents the crowded distance of particle i in the decision space after normalization
  • M represents the dimension of the target space
  • N represents the decision space dimension
  • CD avg, obj ′ and CD avg, dec ′ represent the normalized average crowded distance in the target space and the average crowded distance in the decision space, respectively.
  • V′ i wV i +c 1 r 1 (Pbest i -X i )+c 2 r 2 (Gbest-X i )
  • X i is the current particle position
  • X i ' is the updated particle position
  • V i is the current particle velocity
  • V i ' is the updated particle velocity
  • Pbest is the individual best of each particle
  • Gbest is The population is globally best, it is subGbest;
  • each non-dominated solution set of each population are connected end to end to form a ring topology, and each non-dominated solution set and its adjacent two non-dominated solution sets are regarded as a neighborhood;
  • the first particle in the neighborhood file is regarded as the best Nbest of the neighborhood, and then the global best subGbest of each population is updated.
  • the update method is:
  • V′ i wV i +c 1 r 1 (Pbest i -X i )+c 2 r 2 (Nbest i -X i )
  • the termination condition is generally a specified number of iterations or the obtained solution satisfies a certain error range.
  • the specific steps of S6 are: after the iteration is completed, the non-dominated particles in each non-dominated solution set are Pareto optimal solutions in the decision space, and each Pareto optimal solution is substituted into the objective function to obtain Pareto frontier in target space.
  • the following steps are further included: determining relevant parameters of the particle swarm, randomly generating N particles in the decision space, and initializing the particle swarm.
  • the relevant parameters include the total number of particles N, the particle dimension D, the inertia weight w, the speed control parameter c1 and the speed control parameter c2; the speed and position of each particle are initialized.
  • the present invention provides a multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance, which has the following advantages compared with the prior art:
  • the present invention uses Bayesian adaptive resonance theory to divide the particle swarm. Unlike the previous clustering methods based on Euclidean distance and k-means, this method does not need to pre-set the number of clusters, and uses The particles are clustered unsupervised and adaptively with respect to the posterior probability of all existing clusters;
  • the present invention is based on a multi-swarm mechanism, and different sub-populations simultaneously independently track and save the Pareto optimal solution set of the decision space, which is beneficial to the solution of multi-modal problems; and the present invention also adopts non-dominated sorting and special crowding.
  • the distance sorts the particles of various groups, and uses the global optimal particle swarm algorithm to independently mine each sub-population, which is conducive to discovering the distribution of non-dominated solution sets and Pareto frontiers, which is more conducive to multi-objective multi-mode solution of state problems;
  • each population can obtain guiding information from the adjacent population, so as to realize the communication between the populations. It is beneficial to maintain the diversity of particles, and use the local search based on ring topology to further improve the search efficiency and improve the space exploration ability of the algorithm;
  • Pareto optimal solution set of the decision space and the corresponding Pareto optimal solution of the target space can be output at the same time, and the solutions of multi-objective and multi-modal problems can be presented more intuitively.
  • FIG. 1 accompanying drawing is the overall step flow chart of the multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance provided by the present invention
  • Fig. 2 is the algorithm flow chart of the multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance provided by the present invention
  • FIG. 3 is a schematic diagram of a decision space and a target space in the multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance provided by the present invention
  • FIG. 5 is a schematic diagram of a ring topology formed by a non-dominated solution set in the multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance provided by the present invention.
  • the embodiment of the present invention discloses a multi-objective multi-modal particle swarm optimization method based on Bayesian adaptive resonance.
  • Step 1 Determine the relevant parameters of the particle swarm and initialize the particle swarm.
  • the relevant parameters here mainly include: the total number of particles N, the particle dimension D, the inertia weight w, and the learning factors c1 and c2.
  • the total number of particles N is crucial to the performance of the algorithm. If the number of particles is too small, the diversity of the population will be insufficient, and it will not be enough to fully cover the decision space, resulting in poor decision performance. Conversely, if the population size is too large, it will consume too much computing resources, so it is necessary to choose an appropriate total number of particles. This parameter needs to be set based on actual problem requirements, prior experience, and some preliminary experiments.
  • the particle dimension D is determined by the number of variables. In general, the particle dimension is equal to the number of decision variables.
  • the inertia weight w is the ability of the particle to maintain the motion state of the previous moment.
  • the learning factors c1 and c2 are mainly used to control the degree of influence of the particle by individual cognition and social cognition. Usually, the value of w is 0.7298, and the values of c1 and c2 are both 2.05.
  • the stability and controllability of the aircraft are two very important characteristics of the aircraft. If the stability is good, the force and torque of the aircraft to resist the change of the flight state will be large, the response of the aircraft to the pilot's control will be slow, and the controllability of the aircraft will be reduced. Correspondingly worse. How to reconcile the relationship between the stability and maneuverability of an aircraft is a very worthy trade-off for modern aircraft design. In order to balance the longitudinal stability and maneuverability of the aircraft, it is necessary to determine the position of the aircraft's focus and center of gravity in the aircraft design stage. The focus and center of gravity of the aircraft are two important factors that affect the stability and maneuverability of the aircraft.
  • the stability and maneuverability of the aircraft are regarded as objective functions f 1 , f 2
  • the focus and the position of the center of gravity of the aircraft are regarded as decision variables x 1 , x 2
  • there are 2 decision variables so the The particle dimension N determined in the step is 2.
  • the decision space of the particle is shown on the left side of Figure 3, where "x 1 , x 2 " is the decision variable; it is the space composed of the value range of each dimension variable of the particle; the target space is shown on the right side of Figure 3 , is the space composed of the value range of the objective function determined by the decision space, where "y 1 , y 2 " is the objective function.
  • Randomly generate N particles in the decision space and initialize the speed and position of each particle.
  • N particles are randomly generated within the range of the focus and the center of gravity of the aircraft, and the abscissa x 1 of the particles represents the size of the aircraft.
  • the focus position, the ordinate x 2 of the particle represents the position of the particle's center of gravity.
  • Step 2 Use Bayesian adaptive resonance theory to divide the particle swarm into several populations.
  • Bayesian adaptive resonance is regarded as a clustering method here, and the process is shown in Figure 4.
  • the specific method is as follows:
  • the calculation method of the posterior probability of the particle for the existing cluster is:
  • K represents the number of population existing
  • y l j represents the populations
  • the estimated prior probability of the jth population represents the possibility of population j to particle x;
  • ⁇ j represent the estimated mean vector and covariance matrix of the j population
  • the population sample size is represented by the hypervolume S J , which is the determinant of the Gaussian covariance matrix.
  • the hypervolume is the product of variances in each dimension:
  • d represents the particle dimension, represents variance
  • M J represents the number of particles in the J population
  • I represents the identity matrix
  • N particles can be divided into k populations.
  • Step 3 Sort the particles of each population according to the non-dominated sorting method and the special crowding distance.
  • solution A dominates solution B if one solution A is better than another solution B on all objectives.
  • a solution A is called a non-dominated solution if it is not dominated by other solutions.
  • the target solution of these particles is divided into the first non-dominated layer in the target space.
  • the target solution of these particles is divided into the second non-dominated layer in the target space.
  • All particles in the population are sorted hierarchically according to the above method.
  • the crowding distance is an index of the crowding degree between the target solution of a particle and the target solutions of its adjacent particles in the target space.
  • M represents the number of all objective functions
  • CD im, obj represent the crowding distance of particle i in the dimension of objective function f m (x)
  • the calculation method is:
  • f m (x i+1 ) and f m (x i-1 ) represent the adjacent target solutions of the objective function solution f m ( xi ) of particle i
  • f m (x max ) and f m (x min ) represents the maximum and minimum values of the objective function f m (x).
  • the multi-objective multi-modal problem needs to consider not only the diversity of the target space, but also the diversity of the decision space. Therefore, if the crowded distance of the target space is extended to the decision space, the crowded distance of the decision space CD i can be obtained in the same way, dec ; replace the objective function in the above method with a decision function to obtain the decision space crowding distance CD i, dec ;
  • CD i, obj ′ represents the crowded distance of particle i in the target space after normalization
  • CD i, obj ′ represents the crowded distance of particle i in the decision space after normalization
  • M represents the dimension of the target space
  • N represents the decision space dimension
  • CD avg, obj ′ and CD avg′dec ′ represent the normalized average crowded distance of the target space and the average crowded distance of the decision space, respectively.
  • the non-dominated particles in the first non-dominated layer are sorted in descending order according to the size of the special crowding distance, and the first non-dominated particle has the largest special crowding distance.
  • Step 4 Save the sorted non-dominated solutions of each population in the set of non-dominated solutions of various populations.
  • the non-dominated solution set of each population is established, and the non-dominated particles sorted in descending order according to the special crowding distance are stored in the non-dominated solution set. If a new non-dominated solution is generated subsequently, it will be added to the non-dominated solution set, and the particles dominated by the new non-dominated solution will be deleted.
  • Step 5 The particles of each population are evolved using the global optimal particle swarm optimization algorithm, and the method is as follows:
  • V′ i wV i +c 1 r 1 (Pbest i -X i )+c 2 r 2 (Gbest-X i )
  • X i is the current particle position
  • X i ' is the updated particle position
  • V i is the current particle velocity
  • V i ' is the updated particle velocity
  • Pbest is the individual best of each particle
  • Gbest is The population is globally best, it is subGbest;
  • Step 6 Use the particle swarm optimization algorithm based on ring topology for local search, the method is as follows:
  • each solid circle represents a non-dominated solution set of a population, where "Nset" is a non-dominated solution set.
  • the three non-dominated solution sets included in each dashed line are a neighborhood.
  • V′ i wV i +c 1 r 1 (Pbest i -X i )+c 2 r 2 (Nbesti -X i )
  • Step 7 Repeat the evolution of various groups and the local search process based on ring topology until the termination condition is satisfied.
  • the new global best subGbest obtained in step 6 update the particle velocity and position in the population according to the rules in step 5, and then update the global best according to the rules in step 6, and iterate continuously until the termination condition is satisfied.
  • it is a specified number of iterations or the obtained solution satisfies a certain error range.
  • Step 8 Output the non-dominated particles in each non-dominated solution set and their corresponding Pareto fronts.
  • the non-dominated particles in the non-dominated solution set of each population are the Pareto optimal solutions in the decision space, and they can be brought into the objective function to obtain the Pareto frontier in the objective space.
  • the final output Pareto front is the optimal combination of aircraft stability and maneuverability, and the non-dominated particles in the non-dominated solution set are corresponding to the optimal combination of aircraft stability and maneuverability.
  • the invention can complete the construction of the multi-objective multi-modal particle swarm method based on Bayesian adaptive resonance.
  • the clustering method based on Bayesian adaptive theory is used to divide the particle swarm into multiple sub-populations in the decision space, and then In each population, sort each particle in various groups according to the non-dominated sorting method and the special crowding distance, and store the non-dominated solution into the non-dominated solution set, and the first one is regarded as the global optimum of the population; Then, the particles in the population are updated using the individual optimality of the particles and the global optimality of the population; the non-dominated solution sets are connected end to end to form a closed ring topology, and the particle swarm optimization algorithm based on ring topology is used for local exploration; repeat the above Two update and exploration processes are performed until the termination condition is met, outputting all non-dominated solution sets and Pareto fronts.
  • Using this technology is beneficial to solve the optimization problem of multi-objective and multi-modal functions. It can help people to plan multiple alternative schemes to achieve multiple objectives in reality, provide redundant backup methods, and improve the reliability of engineering practice activities.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Geometry (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Automation & Control Theory (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了基于贝叶斯自适应共振的多目标多模态粒子群优化方法,利用贝叶斯自适应共振理论将所有粒子划分为若干种群;根据非支配排序法和特殊拥挤距离对各个种群的粒子进行排序;利用粒子的个体最优和种群的全局最优对种群中的粒子进行更新;将各种群的非支配解集首尾相连形成一个闭合环形拓扑,利用基于环形拓扑的粒子群优化算法进行局部探索;重复以上两个更新和探索过程直至满足终止条件,输出所有的非支配解集和帕累托前沿。本发明适用于解决多目标多模态问题的优化,既可以在目标空间中找到帕累托前沿的分布,也可以在决策变量空间找到对应的帕累托最优解集,提供冗余备份方法,提高工程实践活动的可靠性。

Description

基于贝叶斯自适应共振的多目标多模态粒子群优化方法 技术领域
本发明涉及优化算法技术领域,更具体的说是涉及基于贝叶斯自适应共振的多目标多模态粒子群优化方法。
背景技术
现实生活中往往存在着互相冲突互相制约的多个优化目标的问题,例如在飞机的设计过程中,既想保证飞机的稳定性,又想追求飞机的操纵性,稳定性和操纵性就是两个互相制约的目标,同时,同一个目标的解决又可能含有多个方案,例如不同的飞机设计方案可能得到相同的稳定性和操纵性,这样的问题被称为多目标多模态问题。其中目标函数所组成的空间被称为目标空间,而用于生成目标函数的变量所组成的空间被称为决策空间。
对于多目标问题,传统的研究一般只注重于寻找目标空间的帕累托前沿,而忽视一个帕累托前沿可能对应决策空间的多个变量组合,除了保证目标空间解的多样性,保证决策空间的多样性对于解决多目标问题也同样重要,因此近些年来人们越来越关注多目标多模态问题的优化方法。
目前,基于自组织映射的智能算法、基于重组策略的进化算法、多模态多目标差分算法、基于环形拓扑的多目标多模态算法和基于聚类的粒子群优化算法等算法被相继提出用于解决这多目标多模态优化问题。由于粒子群优化算法具有高效率和强鲁棒性,近二十年来被广泛的应用于学术界和工程界。对于多模态问题,小生境粒子群算法是一种较好的解决方案;对于多目标问题,基于k-means聚类和基于欧氏距离聚类的粒子群优化算法也被提出过,然而现有的基于聚类的粒子群优化算法往往需要根据解的数量预先设置聚类个数,但是现实中很难提前知道多目标多模态函数的解有多少个,因此过去的 方法一般是取不同个数的聚类数量进行实验,根据实验结果选择较优的聚类数量,这给优化问题带来了很大的不确定性和问题依赖性。
因此,如何提出一种能准确保证目标空间和决策空间的解的多样性的多目标多模态粒子群优化方法是本领域技术人员亟需解决的问题。
发明内容
有鉴于此,本发明提供了基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其目的在于,为了更好的解决多目标多模态优化问题,获取尽可能多的目标空间的帕累托前沿及决策空间的帕累托最优解集,保证目标空间和决策空间的解的多样性。
为了实现上述目的,本发明采用如下技术方案:
基于贝叶斯自适应共振的多目标多模态粒子群优化方法,包括以下步骤:
S1.通过贝叶斯自适应共振理论将粒子群划分为若干个种群;
S2.根据非支配排序法和特殊拥挤距离对各个种群的粒子进行排序:根据非支配排序法在目标空间中对各种群中的所有粒子进行分层排序,对排序后位于第一层非支配层的粒子按照特殊拥挤距离的大小进行降序排序;
S3.将排序后的各个种群的非支配解保存在各种群的非支配解集中;
S4.各个种群的粒子利用基于全局最优粒子群优化算法进化;
S5.利用基于环形拓扑的粒子群优化算法进行局部搜索;
S6.重复各种群的进化和基于环形拓扑的局部搜索过程直至满足终止条件;
S7.输出各个非支配解集中的非支配粒子及相对应的帕累托前沿。
需要说明的是:
粒子的决策空间,即粒子的每一维的取值范围所组成的空间;目标空间,即由决策空间决定的目标函数的取值范围所组成的空间。
在S2中“非支配”,是指:在多目标问题中,若一个解A在所有目标上均优于另一个解B,则解A支配解B;若一个解A没有被其他解支配,则解A称为非支配解。
拥挤距离是指:在目标空间中,粒子的目标解与其相邻粒子的目标解之间拥挤度的指标,拥挤距离越大表明粒子的目标解分布越分散,越能保证解的多样性。
优选的,S1的具体内容包括:
S11.种群选择:为了将不同粒子划分到合适的种群中,计算每个粒子相对所有已存在种群的后验概率最大值,最大后验概率对应的种群作为获胜种群;
粒子对于已存在种群的后验概率的计算方法为:
Figure PCTCN2021070103-appb-000001
Figure PCTCN2021070103-appb-000002
其中,
Figure PCTCN2021070103-appb-000003
代表粒子x对于y j种群的后验概率,K代表已经存在的种群数量,y l代表第j个种群,
Figure PCTCN2021070103-appb-000004
代表第j个种群的估计先验概率,
Figure PCTCN2021070103-appb-000005
代表种群j之于粒子x的可能性;
Figure PCTCN2021070103-appb-000006
通过j种群的多元高斯函数来估计:
Figure PCTCN2021070103-appb-000007
其中,
Figure PCTCN2021070103-appb-000008
和∑ j代表j种群的估计均值向量和协方差矩阵;
根据后验概率最大的种群选择获胜种群j:
Figure PCTCN2021070103-appb-000009
S12.匹配测试:如果获胜种群的样本点容量小于警戒阈值S MAX,则执行S13;否则,根据S11寻找下一个获胜种群;如果所有种群的样本点容量都大 于警戒阈值S MAX,则建立新的种群,新种群的均值向量为粒子本身,协方差矩阵是一个极小的值;
种群样本容量使用超体积S J表示,是高斯协方差矩阵的行列式,对于对角协方差矩阵,超体积为每一维的方差乘积:
Figure PCTCN2021070103-appb-000010
其中,d代表粒子维度,
Figure PCTCN2021070103-appb-000011
代表方差;
S13.学习更新:根据新粒子调整获胜种群的均值向量和协方差矩阵:
Figure PCTCN2021070103-appb-000012
Figure PCTCN2021070103-appb-000013
其中,
Figure PCTCN2021070103-appb-000014
Figure PCTCN2021070103-appb-000015
分别表示加入新粒子后的种群均值向量和协方差,M J表示J种群中的粒子数,I代表单位矩阵。
优选的,非支配排序法包括以下内容:
(4)将粒子代入到多个目标函数中,得到粒子所代表的目标解;
(5)若种群中粒子没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第一层非支配层;
若种群中的粒子除了被第一层非支配层的粒子支配以外,没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第二层非支配层;
按步骤(2)的方法依次对种群中的所有粒子进行分层排序。
优选的,特殊拥挤距离的计算方法如下:
(1)计算目标空间拥挤距离CD i,obj
Figure PCTCN2021070103-appb-000016
其中,M表示所有目标函数的个数,CD im,obj表示粒子i在目标函数f m(x)维度的拥挤距离,计算方法为:
Figure PCTCN2021070103-appb-000017
其中,f m(x i+1)和f m(x i-1)表示粒子i的目标函数解f m(x i)的临近目标解,f m(x max)和f m(x min)表示目标函数f m(x)的最大值和最小值;
同理求得决策空间拥挤距离CD i,dec
(2)对决策空间和目标空间的拥挤距离进行归一化处理;
Figure PCTCN2021070103-appb-000018
Figure PCTCN2021070103-appb-000019
其中,CD i,obj′代表粒子i在归一化后的目标空间拥挤距离,CD i,obj′代表粒子i在归一化后的决策空间拥挤距离,M表示目标空间维度,N表示决策空间维度;
(6)计算粒子i的特殊拥挤距离:
Figure PCTCN2021070103-appb-000020
其中,CD avg,obj′和CD avg,dec′分别代表归一化后的目标空间平均拥挤距离和决策空间的平均拥挤距离。
优选的,S4的具体步骤为:
(1)创建每个粒子的档案用于存储每个粒子的历史信息,将各种群的非支配解集中的排第一位的粒子视为种群全局最佳subGbest,之后种群中的每 个粒子更新自己的速度和位置,更新方法为:
V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Gbest-X i)
X′ i=X i+V i
其中,X i是当前粒子的位置,X i’是更新后粒子的位置,V i是当前粒子的速度,V i’是更新后粒子的速度,Pbest为每个粒子的个体最佳,Gbest为种群全局最佳,则为subGbest;
(2)将更新后粒子存储到粒子档案中,并删除档案中被更新后粒子支配的粒子;
(3)当粒子档案中有能够支配个体最佳Pbest的粒子时,则用能支配个体最佳Pbest的粒子取代个体最佳Pbest成为新的个体最佳,否则不更新个体最佳Pbest;
(4)同理,当个体最佳Pbest中有能够支配全局最佳的粒子subGbest时,则用能够支配全局最佳的粒子subGbest的粒子取代全局最佳subGbest成为新的全局最佳,否则不更新全局最佳subGbest。
优选的,S5的具体步骤为:
(1)将每个种群的非支配解集首尾相连组成一个环形拓扑,把每个非支配解集及与之相邻的两个非支配解集视为一个邻域;
(2)将邻域中的所有非支配粒子放入邻域档案中,根据非支配排序法和特殊拥挤距离对粒子进行排序;
(3)将邻域档案中排第一位的粒子视为邻域最佳Nbest,再更新每个种群的全局最佳subGbest,更新方法为:
V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Nbest i-X i)
X′ i=X i+V i
(4)将每个种群更新后的全局最佳subGbest放入各自的非支配解集中,将非支配解集中的粒子重新根据非支配排序法和特殊拥挤距离进行排序,删除被支配的粒子,将排第一位的粒子作为新的全局最佳subGbest。
优选的,S6的具体步骤为:
根据S5得到的新的种群全局最佳subGbest按照S4中的规则更新种群中的粒子速度和位置,然后再按照S5中的规则更新种群全局最佳,不断循环迭代,直到满足终止条件。
需要说明的是:
终止条件一般为规定的迭代次数或者得到的解满足某一误差范围。
优选的,S6的具体步骤为:迭代完成后,每个非支配解集中的非支配粒子则为决策空间中的帕累托最优解,将各帕累托最优解代入目标函数则得到在目标空间的帕累托前沿。
优选的,利用贝叶斯自适应共振理论将粒子群划分为若干个种群之前还包括以下步骤:确定粒子群的相关参数,在决策空间内随机生成N个粒子,并对粒子群初始化。
优选的,相关参数包括粒子总数量N,粒子维度D,惯性权重w,速度控制参数c1和速度控制参数c2;初始化各个粒子的速度和位置。
经由上述的技术方案可知,本发明公开提供了基于贝叶斯自适应共振的多目标多模态粒子群优化方法,与现有技术相比其优势在于:
首先,本发明利用贝叶斯自适应共振理论对粒子群进行种群划分,不同于以往的基于欧氏距离和基于k-means的聚类方法,该方法不需要预先设定聚类个数,利用粒子相对所有已存在类簇的后验概率,无监督自适应的进行聚类;
其次,本发明基于多种群机制,由不同的子种群同时独立跟踪和保存决策空间的帕累托最优解集,有利于多模态问题的解决;而且本发明还采用非 支配排序和特殊拥挤距离对各种群的粒子进行排序,利用全局最优粒子群算法对每个子种群进行独立的挖掘,有利于发现非支配解集和帕累托前沿的分布,从而更有助于多目标多模态问题的解决;
另外,在得到各子种群的非支配解后,通过找到的各子种群的非支配解之间的交互作用,每个种群都能从相邻种群处获得引导信息,从而实现种群间的交流,有利于保持粒子的多样性,利用基于环拓扑的局部搜索,进一步提高搜索效率,提高算法的空间探索能力;
最后,能够同时输出决策空间的帕累托最优解集和对应的目标空间的帕累托最优解,更直观的将多目标和多模态问题的解呈现出来。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据提供的附图获得其他的附图。
图1附图为本发明提供的基于贝叶斯自适应共振的多目标多模态粒子群优化方法总体步骤流程图;
图2为本发明提供的基于贝叶斯自适应共振的多目标多模态粒子群优化方法算法流程图;
图3为本发明提供的基于贝叶斯自适应共振的多目标多模态粒子群优化方法中决策空间和目标空间的示意图;
图4为本发明提供的基于贝叶斯自适应共振的多目标多模态粒子群优化方法中基于贝叶斯自适应共振的聚类流程图;
图5为本发明提供的基于贝叶斯自适应共振的多目标多模态粒子群优化方法中非支配解集构成的环形拓扑示意图。
具体实施方式
下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
本发明实施例公开了基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其方法步骤如图1所示,算法的具体流程如图2所示,具体内容如下:
步骤一、确定粒子群的相关参数,初始化粒子群。
这里的相关参数主要包括:粒子总数量N,粒子维度D,惯性权重w,学习因子c1和c2。粒子总数N对算法的性能至关重要,如果粒子数量太小,种群的多样性就会不足,不足以充分覆盖决策空间,导致决策性能较差。相反,如果种群规模过大,则会消耗太多的计算资源,因此需要选择合适的粒子总数。需要根据实际问题需求、先前经验和一些初步实验来设置该参数。
粒子维度D由变量的个数决定,一般情况下,粒子维度等于决策变量的个数。惯性权重w即粒子能保持前一时刻运动状态的能力,学习因子c1和c2主要用于控制粒子受个体认知和社会认识的影响程度,通常情况w取值0.7298,c1和c2取值均为2.05。
飞机的稳定性和操纵性是飞机的两个非常重要的特性,稳定性好,飞机抵抗飞行状态变化的力和力矩会很大,飞机对飞行员操纵的响应就会变慢,飞机的操纵性就相应变差。如何协调飞机的稳定性和操纵性之间的关系,对于现代飞机设计是一个非常值得权衡的问题。为了平衡飞行器的纵向稳定性和操纵性,需要在飞机设计阶段确定飞机的焦点和重心的位置,飞机的焦点和重心是影响飞机稳定性和操纵性的两个重要因素。设计飞机时需要在稳定性满足要求的情况下保证操纵性,是一个多目标问题;满足同一稳定性和操 纵性要求的飞机焦点和重心位置又有多种方案,因此这是一个多目标多模态问题,所以本发明提出的技术可以用于解决该飞机设计问题。在一种实施方案中,将飞机的稳定性和操纵性视为目标函数f 1、f 2,将飞机的焦点和重心位置视为决策变量x 1、x 2,决策变量有2个,因此该步骤所确定的粒子维度N即为2。
粒子的决策空间如图3左侧所示,其中“x 1,x 2”即决策变量;是由粒子的每一维变量的取值范围所组成的空间;目标空间如图3右侧所示,是由决策空间决定的目标函数的取值范围所组成的空间,其中“y 1,y 2”即目标函数。
在决策空间内随机生成N个粒子,初始化各个粒子的速度和位置。
在一种实施方案中,根据飞机的焦点和重心位置的取值范围,随机地在两者的取值范围内生成N个粒子(x 1,x 2),粒子的横坐标x 1代表飞机的焦点位置,粒子的纵坐标x 2代表粒子的重心位置。
步骤二:利用贝叶斯自适应共振理论将粒子群划分为若干个种群。
贝叶斯自适应共振在此视为一种聚类方法,流程如图4所示,具体方法如下:
(1)种群选择:为了将不同粒子划分到合适的种群中,计算每个粒子相对所有已存在类簇的后验概率最大值,最大后验概率对应的种群作为获胜种群。
粒子对于已存在类簇的后验概率的计算方法为:
Figure PCTCN2021070103-appb-000021
Figure PCTCN2021070103-appb-000022
其中,
Figure PCTCN2021070103-appb-000023
代表粒子x对于y j种群的后验概率,K代表已经存在的种群数量,y l代表第j个种群,
Figure PCTCN2021070103-appb-000024
代表第j个种群的估计先验概率,
Figure PCTCN2021070103-appb-000025
代表种群j之于粒子x的可能性;
Figure PCTCN2021070103-appb-000026
通过j种群的多元高斯函数来估计:
Figure PCTCN2021070103-appb-000027
其中,
Figure PCTCN2021070103-appb-000028
和∑ j代表j种群的估计均值向量和协方差矩阵;
根据后验概率最大的种群选择获胜种群j:
Figure PCTCN2021070103-appb-000029
(2)匹配测试:如果获胜种群的样本点容量小于警戒阈值S MAX,则执行S13;否则,根据S11寻找下一个获胜种群;如果所有种群的样本点容量都大于警戒阈值S MAX,则建立新的种群,新种群的均值向量为粒子本身,协方差矩阵是一个极小值;
种群样本容量使用超体积S J表示,是高斯协方差矩阵的行列式,对于对角协方差矩阵,超体积为每一维的方差乘积:
Figure PCTCN2021070103-appb-000030
其中,d代表粒子维度,
Figure PCTCN2021070103-appb-000031
代表方差;
(3)学习更新:根据新粒子调整获胜种群的均值向量和协方差矩阵:
Figure PCTCN2021070103-appb-000032
Figure PCTCN2021070103-appb-000033
其中,
Figure PCTCN2021070103-appb-000034
Figure PCTCN2021070103-appb-000035
分别表示加入新粒子后的种群均值向量和协方差,M J表示J种群中的粒子数,I代表单位矩阵。
使用该方法可以将N个粒子分成k个种群。
步骤三:根据非支配排序法和特殊拥挤距离将各个种群的粒子进行排序。
1.非支配排序法的解释如下:
在多目标问题中,若一个解A在所有目标上均优于另一个解B,则解A支配解B。若一个解A没有被其他解支配,则解A称为非支配解。
在一种实施方案中,认为飞机的稳定性和操纵性均越大越好,若一个方案A的飞机焦点和重心的位置所决定的飞机的稳定性和操纵性均大于另一个方案B的稳定性和操纵性,则方案A支配方案B。
将粒子带入到多个目标函数中,即得到该粒子所代表的目标解。
如果种群中粒子没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第一层非支配层。
如果种群中的粒子除了被第一次非支配层的粒子支配以外,没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第二层非支配层。
按以上方法依次对种群中的所有粒子进行分层排序。
2.特殊拥挤距离的解释如下:
拥挤距离是指在目标空间中,粒子的目标解与其相邻粒子的目标解之间拥挤度的指标,拥挤距离越大表明粒子的目标解分布越分散,越能保证解的多样性。计算目标空间拥挤距离CD i,obj
Figure PCTCN2021070103-appb-000036
其中,M表示所有目标函数的个数,CD im,obj表示粒子i在目标函数f m(x)维度的拥挤距离,计算方法为:
Figure PCTCN2021070103-appb-000037
其中,f m(x i+1)和f m(x i-1)表示粒子i的目标函数解f m(x i)的临近目标解,f m(x max)和f m(x min)表示目标函数f m(x)的最大值和最小值。
多目标多模态问题不仅需要考虑目标空间的多样性,还需要考虑决策空间的多样性,因此将目标空间的拥挤距离延伸到决策空间中,则同理可求得决策空间拥挤距离CD i,dec;即将上述方法中的目标函数替换为决策函数即可求得决策空间拥挤距离CD i,dec
为了比较决策空间和目标空间的拥挤距离,需要对决策空间和目标空间的拥挤距离进行归一化处理;
Figure PCTCN2021070103-appb-000038
Figure PCTCN2021070103-appb-000039
其中,CD i,obj′代表粒子i在归一化后的目标空间拥挤距离,CD i,obj′代表粒子i在归一化后的决策空间拥挤距离,M表示目标空间维度,N表示决策空间维度;
(7)计算粒子i的特殊拥挤距离:
Figure PCTCN2021070103-appb-000040
其中,CD avg,obj′和CD avg′dec′分别代表归一化后的目标空间平均拥挤距离和决策空间的平均拥挤距离。
最后对非支配排序后的位于第一层非支配层的粒子按照特殊拥挤距离的大小进行降序排序,排第一位的非支配粒子拥有最大的特殊拥挤距离。
步骤四:将排序后的各个种群的非支配解保存在各种群的非支配解集中。
建立各个种群的非支配解集,将各种群按照特殊拥挤距离降序排序的非支配粒子保存在非支配解集中。后续如果产生新的非支配解,则将其加入非支配解集,删除被新的非支配解支配的粒子。
步骤五:各个种群的粒子利用基于全局最优粒子群优化算法进化,其方法如下:
(1)创建每个粒子的档案用于存储每个粒子的历史信息,将各种群的非支配解集中的排第一位的粒子视为种群全局最佳subGbest,之后种群中的每个粒子根据以下公式更新自己的速度和位置:
V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Gbest-X i)
X′ i=X i+V i
其中,X i是当前粒子的位置,X i’是更新后粒子的位置,V i是当前粒子的速度,V i’是更新后粒子的速度,Pbest为每个粒子的个体最佳,Gbest为种群全局最佳,则为subGbest;
(2)将更新后粒子存储到粒子档案中,并删除档案中被更新后粒子支配的粒子;
(3)当粒子档案中有能够支配个体最佳Pbest的粒子时,则用能支配个体最佳Pbest的粒子取代个体最佳Pbest成为新的个体最佳,否则不更新个体最佳Pbest;
(4)同理,当个体最佳Pbest中有能够支配全局最佳的粒子subGbest时,则用能够支配全局最佳的粒子subGbest的粒子取代全局最佳subGbest成为新的全局最佳,否则不更新全局最佳subGbest。
步骤六:利用基于环形拓扑的粒子群优化算法进行局部搜索,其方法如下:
(1)将每个种群的非支配解集首尾相连组成一个环形拓扑,如图5所示,每个实线圆圈代表一个种群的非支配解集,其中“Nset”即为一个非支配解集,每个虚线所包含的三个非支配解集即为一个邻域。
(2)将邻域中的所有非支配粒子放入邻域档案中,根据非支配排序法和特殊拥挤距离对粒子进行排序。
(3)将邻域档案中排第一位的粒子视为邻域最佳Nbest,之后根据以下公式更新每个种群的全局最佳subGbest:
V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Nbesti-X i)
X′ i=X i+V i
(4)之后将每个种群更新后的全局最佳subGbest放入各自的非支配解集中,将非支配解集中的粒子重新根据非支配排序法和特殊拥挤距离进行排序,删除被支配的粒子,将排第一位的粒子作为新的全局最佳subGbest。
步骤七:重复各种群的进化和基于环形拓扑的局部搜索过程直至满足终止条件。
根据步骤六得到的新全局最佳subGbest按照步骤五中的规则更新种群中的粒子速度和位置,然后再按照步骤6中的规则更新全局最佳,不断的循环迭代,直到满足终止条件,终止条件一般为规定的迭代次数或者得到的解满足某一误差范围。
步骤八:输出各个非支配解集中的非支配粒子及其对应的帕累托前沿。
迭代完成后,每个种群非支配解集中的非支配粒子即为决策空间中的帕累托最优解,将它们带入目标函数即可得到在目标空间的帕累托前沿。
在一种实施方案中,最终输出帕累托前沿即为飞机的稳定性和操纵性的最优组合,非支配解集中的非支配粒子即为飞机稳定性和操纵性的最优组合所对应的飞机焦点和重心的最优位置。
本发明可以完成对基于贝叶斯自适应共振的多目标多模态粒子群方法的构建,首先利用基于贝叶斯自适应理论的聚类方法在决策空间将粒子群划分为多个子种群,而后在每个种群中根据非支配排序法和特殊拥挤距离对各种群中的每个粒子进行排序并将非支配解存入非支配解集中,排序第一位的作为该种群的全局最优;然后利用粒子的个体最优和种群的全局最优对种群中的粒子进行更新;再将非支配解集首尾相连形成一个闭合环形拓扑,利用基于环形拓扑的粒子群优化算法进行局部探索;重复以上两个更新和探索过程直至满足终止条件,输出所有的非支配解集和帕累托前沿。
使用这一技术有利于解决多目标多模态函数的优化问题,可以在现实中帮助人们规划多个可选择的方案从而实现多个目标,提供冗余备份方法,提高工程实践活动的可靠性。
本说明书中各个实施例采用递进的方式描述,每个实施例重点说明的都是与其他实施例的不同之处,各个实施例之间相同相似部分互相参见即可。对于实施例公开的装置而言,由于其与实施例公开的方法相对应,所以描述的比较简单,相关之处参见方法部分说明即可。
对所公开的实施例的上述说明,使本领域专业技术人员能够实现或使用本发明。对这些实施例的多种修改对本领域的专业技术人员来说将是显而易见的,本文中所定义的一般原理可以在不脱离本发明的精神或范围的情况下,在其它实施例中实现。因此,本发明将不会被限制于本文所示的这些实施例,而是要符合与本文所公开的原理和新颖特点相一致的最宽的范围。

Claims (10)

  1. 基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其特征在于,包括以下步骤:
    S1.通过贝叶斯自适应共振理论将粒子群划分为若干个种群;
    S2.根据非支配排序法和特殊拥挤距离对各个种群的粒子进行排序:根据非支配排序法在目标空间中对各种群中的所有粒子进行分层排序,对排序后位于第一层非支配层的粒子按照特殊拥挤距离的大小进行降序排序;
    S3.将排序后的各个种群的非支配解保存在各种群的非支配解集中;
    S4.各个种群的粒子利用基于全局最优粒子群优化算法进化;
    S5.利用基于环形拓扑的粒子群优化算法进行局部搜索;
    S6.重复各种群的进化和基于环形拓扑的局部搜索过程直至满足终止条件;
    S7.输出各个非支配解集中的非支配粒子及相对应的帕累托前沿。
  2. 根据权利要求1所述的基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其特征在于,S1的具体内容包括:
    S11.种群选择:为了将不同粒子划分到合适的种群中,计算每个粒子相对所有已存在种群的后验概率最大值,最大后验概率对应的种群作为获胜种群;
    粒子对于已存在种群的后验概率的计算方法为:
    Figure PCTCN2021070103-appb-100001
    Figure PCTCN2021070103-appb-100002
    其中,
    Figure PCTCN2021070103-appb-100003
    代表粒子x对于y j种群的后验概率,K代表已经存在的种群数量,y l代表第j个种群,
    Figure PCTCN2021070103-appb-100004
    代表第j个种群的估计先验概率,
    Figure PCTCN2021070103-appb-100005
    代表种群j之于粒子x的可能性;
    Figure PCTCN2021070103-appb-100006
    通过j种群的多元高斯函数来估计:
    Figure PCTCN2021070103-appb-100007
    其中,
    Figure PCTCN2021070103-appb-100008
    和∑ j代表j种群的估计均值向量和协方差矩阵;
    根据后验概率最大的种群选择获胜种群j:
    Figure PCTCN2021070103-appb-100009
    S12.匹配测试:如果获胜种群的样本点容量小于警戒阈值S MAX,则执行S13;否则,根据S11寻找下一个获胜种群;如果所有种群的样本点容量都大于警戒阈值S MAX,则建立新的种群,新种群的均值向量为粒子本身;
    种群样本容量使用超体积S J表示,是高斯协方差矩阵的行列式,对于对角协方差矩阵,超体积为每一维的方差乘积:
    Figure PCTCN2021070103-appb-100010
    其中,d代表粒子维度,
    Figure PCTCN2021070103-appb-100011
    代表方差;
    S13.学习更新:根据新粒子调整获胜种群的均值向量和协方差矩阵:
    Figure PCTCN2021070103-appb-100012
    Figure PCTCN2021070103-appb-100013
    其中,
    Figure PCTCN2021070103-appb-100014
    Figure PCTCN2021070103-appb-100015
    分别表示加入新粒子后的种群均值向量和协方差,M J表示J种群中的粒子数,I代表单位矩阵。
  3. 根据权利要求1所述的基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其特征在于,非支配排序法包括以下内容:
    (1)将粒子代入到多个目标函数中,得到粒子所代表的目标解;
    (2)若种群中粒子没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第一层非支配层;
    若种群中的粒子除了被第一层非支配层的粒子支配以外,没有被其他粒子所支配,则在目标空间中将这些粒子的目标解划分为第二层非支配层;
    按步骤(2)的方法依次对种群中的所有粒子进行分层排序。
  4. 根据权利要求1所述的基于贝叶斯自适应共振的多目标多模态粒子群优化方法,其特征在于,特殊拥挤距离的计算方法如下:
    (1)计算目标空间拥挤距离CD i,obj
    Figure PCTCN2021070103-appb-100016
    其中,M表示所有目标函数的个数,CD im,obj表示粒子i在目标函数f m(x)维度的拥挤距离,计算方法为:
    Figure PCTCN2021070103-appb-100017
    其中,f m(x i+1)和f m(x i-1)表示粒子i的目标函数解f m(x i)的临近目标解,f m(x max)和f m(x min)表示目标函数f m(x)的最大值和最小值;
    同理求得决策空间拥挤距离CD i,dec
    (2)对目标空间和决策空间的拥挤距离进行归一化处理;
    Figure PCTCN2021070103-appb-100018
    Figure PCTCN2021070103-appb-100019
    其中,CD i,obj′代表粒子i在归一化后的目标空间拥挤距离,CD i,obj′代表粒子i在归一化后的决策空间拥挤距离,M表示目标空间维度,N表示决策空间维度;
    (3)计算粒子i的特殊拥挤距离:
    Figure PCTCN2021070103-appb-100020
    其中,CD avg,obj′和CD avg,dsc′分别代表归一化后的目标空间平均拥挤距离和决策空间的平均拥挤距离。
  5. 根据权利要求1所述的一种多目标多模态离子群优化方法,其特征在于,S4的具体步骤为:
    (1)创建每个粒子的档案用于存储每个粒子的历史信息,将各种群的非支配解集中的排第一位的粒子视为种群全局最佳subGbest,之后种群中的每个粒子更新自己的速度和位置,更新方法为:
    V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Gbest-X i)
    X′ i=X i+V i
    其中,X i是当前粒子的位置,X i’是更新后粒子的位置,V i是当前粒子的速度,V i’是更新后粒子的速度,Pbest为每个粒子的个体最佳,Gbest为种群全局最佳,则为subGbest;
    (2)将更新后粒子存储到粒子档案中,并删除档案中被更新后粒子支配的粒子;
    (3)当粒子档案中有能够支配个体最佳Pbest的粒子时,则用能支配个体最佳Pbest的粒子取代个体最佳Pbest成为新的个体最佳,否则不更新个体最佳Pbest;
    (4)同理,当个体最佳Pbest中有能够支配全局最佳的粒子subGbest时,则用能够支配全局最佳的粒子subGbest的粒子取代全局最佳subGbest成为新的全局最佳,否则不更新全局最佳subGbest。
  6. 根据权利要求1所述的一种多目标多模态离子群优化方法,其特征在于,S5的具体步骤为:
    (1)将每个种群的非支配解集首尾相连组成一个环形拓扑,把每个非支配解集及与之相邻的两个非支配解集视为一个邻域;
    (2)将邻域中的所有非支配粒子放入邻域档案中,根据非支配排序法和特殊拥挤距离对粒子进行排序;
    (3)将邻域档案中排第一位的粒子视为邻域最佳Nbest,再更新每个种群的全局最佳subGbest,更新方法为:
    V′ i=wV i+c 1r 1(Pbest i-X i)+c 2r 2(Nbest i-X i)
    X′ i=X i+V i
    (4)将每个种群更新后的全局最佳subGbest放入各自的非支配解集中,将非支配解集中的粒子重新根据非支配排序法和特殊拥挤距离进行排序,删除被支配的粒子,将排第一位的粒子作为新的全局最佳subGbest。
  7. 根据权利要求1所述的一种多目标多模态离子群优化方法,其特征在于,S6的具体步骤为:
    根据S5得到的新的种群全局最佳subGbest按照S4中的规则更新种群中的粒子速度和位置,然后再按照S5中的规则更新种群全局最佳,不断循环迭代,直到满足终止条件。
  8. 根据权利要求1所述的一种多目标多模态离子群优化方法,其特征在于,S6的具体步骤为:迭代完成后,每个非支配解集中的非支配粒子则为决策空间中的帕累托最优解,将各帕累托最优解代入目标函数则得到在目标空间的帕累托前沿。
  9. 根据权利要求1所述的一种多目标多模态离子群优化方法,其特征在于,利用贝叶斯自适应共振理论将粒子群划分为若干个种群之前还包括以下步 骤:确定粒子群的相关参数,在决策空间内随机生成N个粒子,并对粒子群初始化。
  10. 根据权利要求9所述的一种多目标多模态离子群优化方法,其特征在于,相关参数包括粒子总数量N,粒子维度D,惯性权重w,速度控制参数c1和速度控制参数c2;初始化各个粒子的速度和位置。
PCT/CN2021/070103 2020-07-08 2021-01-04 基于贝叶斯自适应共振的多目标多模态粒子群优化方法 WO2022007376A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010652340.8A CN111814251B (zh) 2020-07-08 2020-07-08 基于贝叶斯自适应共振的多目标多模态粒子群优化方法
CN202010652340.8 2020-07-08

Publications (1)

Publication Number Publication Date
WO2022007376A1 true WO2022007376A1 (zh) 2022-01-13

Family

ID=72841956

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/070103 WO2022007376A1 (zh) 2020-07-08 2021-01-04 基于贝叶斯自适应共振的多目标多模态粒子群优化方法

Country Status (2)

Country Link
CN (1) CN111814251B (zh)
WO (1) WO2022007376A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113848970A (zh) * 2021-08-31 2021-12-28 中南大学 一种车辆-无人机多目标协同路径规划方法
CN114779467A (zh) * 2022-04-27 2022-07-22 吉林大学 一种基于探测器特性的新型光谱仪膜系组合的选择方法
CN115310353A (zh) * 2022-07-26 2022-11-08 明珠电气股份有限公司 一种基于快速多目标优化的电力变压器设计方法
CN115357019A (zh) * 2022-08-23 2022-11-18 福州大学 一种无人船多参数粒子群辨识系统及方法
CN116091046A (zh) * 2023-04-06 2023-05-09 北京理工大学 基于分阶段启发式算法的装备群多波次任务维修规划方法
CN116400318A (zh) * 2023-06-08 2023-07-07 中国人民解放军国防科技大学 基于在线粒子群优化的多观测目标位置估计方法及装置
CN116518982A (zh) * 2023-07-03 2023-08-01 江西啄木蜂科技有限公司 一种低空林业监测遥感无人机路径多目标规划方法
CN116629085A (zh) * 2023-06-15 2023-08-22 中国建筑第二工程局有限公司 基于粒子群算法及卷积神经网络的防水涂料最优化方法
CN116757243A (zh) * 2023-06-27 2023-09-15 哈尔滨工程大学 基于混沌Levy多目标烟花算法的多任务分配方法
CN117060408A (zh) * 2023-10-13 2023-11-14 国网浙江省电力有限公司宁波供电公司 新能源发电预测方法及系统
CN117575299A (zh) * 2024-01-17 2024-02-20 南京信息工程大学 一种基于改进粒子群算法的协同多任务分配方法
CN117786824A (zh) * 2024-02-27 2024-03-29 华东交通大学 一种基于多目标优化的隧道环境参数设计方法及系统

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814251B (zh) * 2020-07-08 2021-03-26 北京航空航天大学 基于贝叶斯自适应共振的多目标多模态粒子群优化方法
CN112417770B (zh) * 2020-12-09 2022-03-18 浙江工业大学 一种基于多模态多目标粒子群优化算法的选址优化方法
CN113343380B (zh) * 2021-05-31 2022-09-09 温州大学 基于多目标粒子群算法的强迫风冷散热器优化方法及系统
CN114363464B (zh) * 2021-12-30 2023-04-21 华南理工大学 一种遏制诈骗信息传播的方法及其系统
CN114553489B (zh) * 2022-01-24 2023-05-26 北京北卡星科技有限公司 一种基于多目标优化算法的工业控制系统安全防护方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106203689A (zh) * 2016-07-04 2016-12-07 大连理工大学 一种梯级水电站群联合运行多目标优化调度方法
CN108268979A (zh) * 2018-02-01 2018-07-10 北京科技大学 一种基于演化模糊关联规则的中厚板质量预测方法
CN111077896A (zh) * 2019-12-27 2020-04-28 南京理工大学 一种基于改进分层算法的充液挠性航天器参数优化方法
CN111814251A (zh) * 2020-07-08 2020-10-23 北京航空航天大学 基于贝叶斯自适应共振的多目标多模态粒子群优化方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101763528A (zh) * 2009-12-25 2010-06-30 深圳大学 基于贝叶斯网络的基因调控网络构建方法
CN108846411A (zh) * 2018-05-03 2018-11-20 武汉理工大学 基于配对稀疏贝叶斯极限学习机的主减速器故障诊断方法
WO2020000248A1 (zh) * 2018-06-27 2020-01-02 大连理工大学 一种基于空间重构的航空发动机过渡态加速过程关键性能参数预测方法
CN111144581A (zh) * 2019-12-31 2020-05-12 杭州雅拓信息技术有限公司 一种机器学习超参数调节方法及系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106203689A (zh) * 2016-07-04 2016-12-07 大连理工大学 一种梯级水电站群联合运行多目标优化调度方法
CN108268979A (zh) * 2018-02-01 2018-07-10 北京科技大学 一种基于演化模糊关联规则的中厚板质量预测方法
CN111077896A (zh) * 2019-12-27 2020-04-28 南京理工大学 一种基于改进分层算法的充液挠性航天器参数优化方法
CN111814251A (zh) * 2020-07-08 2020-10-23 北京航空航天大学 基于贝叶斯自适应共振的多目标多模态粒子群优化方法

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113848970A (zh) * 2021-08-31 2021-12-28 中南大学 一种车辆-无人机多目标协同路径规划方法
CN114779467A (zh) * 2022-04-27 2022-07-22 吉林大学 一种基于探测器特性的新型光谱仪膜系组合的选择方法
CN115310353A (zh) * 2022-07-26 2022-11-08 明珠电气股份有限公司 一种基于快速多目标优化的电力变压器设计方法
CN115310353B (zh) * 2022-07-26 2024-02-20 明珠电气股份有限公司 一种基于快速多目标优化的电力变压器设计方法
CN115357019A (zh) * 2022-08-23 2022-11-18 福州大学 一种无人船多参数粒子群辨识系统及方法
CN116091046A (zh) * 2023-04-06 2023-05-09 北京理工大学 基于分阶段启发式算法的装备群多波次任务维修规划方法
CN116400318A (zh) * 2023-06-08 2023-07-07 中国人民解放军国防科技大学 基于在线粒子群优化的多观测目标位置估计方法及装置
CN116400318B (zh) * 2023-06-08 2023-08-04 中国人民解放军国防科技大学 基于在线粒子群优化的多观测目标位置估计方法及装置
CN116629085A (zh) * 2023-06-15 2023-08-22 中国建筑第二工程局有限公司 基于粒子群算法及卷积神经网络的防水涂料最优化方法
CN116757243A (zh) * 2023-06-27 2023-09-15 哈尔滨工程大学 基于混沌Levy多目标烟花算法的多任务分配方法
CN116757243B (zh) * 2023-06-27 2024-01-23 哈尔滨工程大学 基于混沌Levy多目标烟花算法的多任务分配方法
CN116518982B (zh) * 2023-07-03 2023-09-19 江西啄木蜂科技有限公司 一种低空林业监测遥感无人机路径多目标规划方法
CN116518982A (zh) * 2023-07-03 2023-08-01 江西啄木蜂科技有限公司 一种低空林业监测遥感无人机路径多目标规划方法
CN117060408A (zh) * 2023-10-13 2023-11-14 国网浙江省电力有限公司宁波供电公司 新能源发电预测方法及系统
CN117060408B (zh) * 2023-10-13 2024-05-28 国网浙江省电力有限公司宁波供电公司 新能源发电预测方法及系统
CN117575299A (zh) * 2024-01-17 2024-02-20 南京信息工程大学 一种基于改进粒子群算法的协同多任务分配方法
CN117575299B (zh) * 2024-01-17 2024-04-02 南京信息工程大学 一种基于改进粒子群算法的协同多任务分配方法
CN117786824A (zh) * 2024-02-27 2024-03-29 华东交通大学 一种基于多目标优化的隧道环境参数设计方法及系统
CN117786824B (zh) * 2024-02-27 2024-06-04 华东交通大学 一种基于多目标优化的隧道环境参数设计方法及系统

Also Published As

Publication number Publication date
CN111814251A (zh) 2020-10-23
CN111814251B (zh) 2021-03-26

Similar Documents

Publication Publication Date Title
WO2022007376A1 (zh) 基于贝叶斯自适应共振的多目标多模态粒子群优化方法
Katuwal et al. Enhancing multi-class classification of random forest using random vector functional neural network and oblique decision surfaces
CN111553469A (zh) 一种无线传感器网络数据融合方法、装置和存储介质
CN106991442A (zh) 混合蛙跳算法的自适应核k‑means方法与系统
CN101740029B (zh) 应用于基于矢量量化的说话人识别的三粒子协同优化方法
CN116402002B (zh) 一种用于芯片布局问题的多目标分层强化学习方法
CN111625998A (zh) 一种层叠太阳能电池结构优化方法
CN113255873A (zh) 一种聚类天牛群优化方法、系统、计算机设备和存储介质
Han et al. Locating multiple equivalent feature subsets in feature selection for imbalanced classification
Li et al. Dual-stage hybrid learning particle swarm optimization algorithm for global optimization problems
CN110928873B (zh) 基于改进非支配排序遗传算法的空间对象索引与查询方法
Yan et al. A data clustering algorithm based on mussels wandering optimization
Nayak et al. A modified differential evolution-based fuzzy multi-objective approach for clustering
Tang et al. Using improved glowworm swarm optimization algorithm for clustering analysis
Luo et al. An entropy driven multiobjective particle swarm optimization algorithm for feature selection
Tareq et al. A new density-based method for clustering data stream using genetic algorithm
CN109063779A (zh) 一种云制造资源集群k-means聚类方法
Lemaire et al. An initialization scheme for supervized k-means
Antonelli et al. Evolutionary fuzzy classifiers for imbalanced datasets: an experimental comparison
Roy et al. A modified brainstorm optimization for clustering using hard c-means
Qian et al. Data mining algorithm based on feature weighting
Prabhat A genetic-fuzzy algorithm to discover fuzzy classification rules for mixed attributes datasets
CN110321938B (zh) 一种智能无人集群的状态空间构建方法及装置
de Lima et al. Ensembles of evolutionary extreme learning machines through differential evolution and fitness sharing
Tang et al. A review of glowworm swarm optimization algorithm

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21837475

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21837475

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 23.05.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 21837475

Country of ref document: EP

Kind code of ref document: A1