48
Multiobjective and Many-Objective Optimization using Evolutionary Algorithms 2019-4-16 1 Lin Qiuzhen College of Computer Science and Software Engineering, Shenzhen University, Shenzhen [email protected]

Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

  • Upload
    others

  • View
    20

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Multiobjective and Many-Objective Optimization using Evolutionary Algorithms

2019-4-16 1

Lin Qiuzhen College of Computer Science and Software Engineering,

Shenzhen University, Shenzhen

[email protected]

Page 2: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 2

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble MOEA Framework

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 3: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 3

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble MOEA Framework

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 4: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Multi/Many-objective Optimization Problem (MOP/MaOP)

2019-4-16 4

1 2Minimize ( ) ( ( ), ( ),..., ( ))subject to :

mF x f x f x f xx

m objective vectors

search space

Page 5: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Dominance & Pareto Optimal Solutions

2019-4-16 5

y dominates x if and only if:Ø y is no worse than x in any obj, and Ø y is better than x in at least one obj.For examples:• B dominates C• A and C are not comparable

x is Pareto optimal if no other solution dominates it.

Pareto set (PS) = the set of all Pareto optimal solutions in the decision space (x-space).

Pareto front (PF) = the image of the PS in the objective space (F-space).

Page 6: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 6

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble MOEA Framework

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 7: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Multiobjective Evolutionary Algorithms (MOEAs)

2019-4-16 7

Ø Approximate the PS (PF) in a single run

Ø Main procedures:

Mating Selection:

select parent pairs to constitute a mating pool

Reproduction:

create offspring based on evolutionary operators

Environmental Selection:

preserve a set of non-dominated solutions to approximate the PS (PF)

Ø Three typical MOEAs: Pareto-dominance-based, indicator-based and decomposition-based MOEAs.

General framework of MOEAs

Page 8: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Pareto-dominance-based MOEAs

2019-4-16 8

The most famous MOEA: NSGA-II [1]

Other MOEAs:SPEA2,PAES2,

…NSGA-III,

VaEAθ-DEA

[1] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput., vol. 6, no. 2, pp. 182–197, 2002.

Page 9: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Pareto-dominance-based MOEAs(need to consider diversity)

2019-4-16 9

Non-dominated sorting in NSGA-II

Page 10: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Indicator-based MOEAs

2019-4-16 10

Ø Selection based on performance indicator (hypervolume,inverted generational distance, Hausdorff distance, etc.)

Ø Relevant algorithms:IBEA, SMS-EMOA, HyPE,…

Ø Due to the high computational cost, this type of MOEAs is less considered than others, especially for MaOPs.

Page 11: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Decomposition-based MOEAs

2019-4-16 11

Ø Decompose a MOP into multiple single optimization problems or multiple simple MOPs

Ø A set of weight vectors to determine diversity and convergence measured by weighted aggregation objective values

Ø Relevant algorithms: MOEA/D, MOEA/D-M2M, MOEA/D-STM, MOEA/D-IR,

[2] Q. F. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm based on decomposition,” IEEE Trans. Evol. Comput., vol. 11 , no. 6, pp. 712–731, 2007.

Page 12: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Decomposition-based MOEAs

2019-4-16 12

Ø Three common aggregation approaches: Weighted sum approach (WS), Tchbycheff approach (TCH) and penalty-based boundary intersection approach (PBI)

(a) WS, (b) TCH, (c) PBIPF: Pareto frontIR: improvement region

Page 13: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 13

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble Framework for MOPs

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 14: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

An effective ensemble framework for MOPs

2019-4-16 14

Ø Motivation: Make full use of the advantages of three typical MOEAsØ Basic ideal: by combining the advantages of various evolutionary operators

and selection criteria that are run on multiple populations.

• 50 offspring solutions get by SBX and DE • solutions from SBX arecentralized around four corner points • solutions from DE are distributed more evenly

Page 15: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

An effective ensemble framework for MOPs

2019-4-16 15

Ø Motivation: Make full use of the advantages of three typical MOEAsØ Basic ideal: by combining the advantages of various evolutionary operators

and selection criteria that run on multiple populations.

• decomposition-based selection criterion with uniform weight vectors is not so good at tackling the problems with discontinuous and irregular PFs

• Pareto-based selection criterion is unable to provide strong convergence pressure on UF2

Page 16: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 16

• Two mechanisms, namely competition and cooperation, are employed to drive the running of the ensembles

• Competition on evolutionary operators

• Cooperation on selection criteria

• A simple algorithm based on the idea of ensemble framework is introduced by employing Pareto and decomposition-based populations, termed EF-PD

An effective ensemble framework for MOPs

Page 17: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

• EF-PDCompetition on SBX and DE

Cooperation on Pareto-based anddecomposition-based criterion

PP indicates the population evolvedby SBX and selected by Pareto-based criterion.DP denotes the population evolvedby DE and selected by decomposition-based criterion

2019-4-16 17

An effective ensemble framework for MOPs

Page 18: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Competition on SBX and DEBoth SBX and DE are adaptively run according to their credits (FSBX and FDE)

where at generation g and denote the index sets of subproblems enhanced by SBX and DE. and are the number of executions of SBX and DE indicates the enhancement of the i-th subproblem by TCH function, as:

the enhancement brought by the new solution y associated to the i-th subproblem over the original associated solution x

2019-4-16 18

EF-PD

gSBXI g

DEIgSBXN g

DENif

Page 19: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Competition on SBX and DEThe normalized credits (FS and FD) can be obtained by

Based on FS and FD, the number of executions of DE and SBX at generation g+1 can be calculated by

2019-4-16 19

EF-PD

Page 20: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Cooperation on Pareto-based and decomposition-based criterion

SDE (the offspring set generated by DE), SSBX (the offspring set generated by SBX)2019-4-16 20

EF-PD

Page 21: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 21

EF-PDCooperation on Pareto-based and decomposition-based criterion

and respectively denote the selectionmechanisms associated to the use of the decomposition-based and Pareto-based approaches

Page 22: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

• Experimental results on ZDT and WFG problems

2019-4-16 22

EF-PD

Performance indicator: Inverted Generational Distance (IGD)

Page 23: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

• Experimental results on DTLZ and UF problems

2019-4-16 23

EF-PD

Performance indicator: Inverted Generational Distance (IGD)

Page 24: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 24

EF-PD

Page 25: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 25

EF-PD

Page 26: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 26

EF-PD

Page 27: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

[3] W. Wang, S. Yang, Q, Lin, Q. Zhang, K. Wong, C. A. Coello Coello, J. Chen, “An Effective Ensemble Framework for Multi-objective Optimization,” IEEE Transactions on Evolutionary Computation, in press, 2018.

2019-4-16 27

A generalized ensemble framework for MOPs

Page 28: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 28

A generalized ensemble framework for MOPs

[3] W. Wang, S. Yang, Q, Lin, Q. Zhang, K. Wong, C. A. Coello Coello, J. Chen, “An Effective Ensemble Framework for Multi-objective Optimization,” IEEE Transactions on Evolutionary Computation, in press, 2018.

Page 29: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 29

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble MOEA Framework

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 30: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Many-objective evolutionary algorithms (MaOEAs)

2019-4-16 30

The major issue for solving MaOPs

inefficiency of Pareto

dominance

difficulty of diversity

maintenance

high computation

al cost

inefficiency of variation operators

Pareto-dominance-based

MOEAs

Decomposition-based MOEAs

Indicator-based MOEAs For all MOEAs

Page 31: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 31

Main ideal:Ø Without considering uniformly distributed weight vectors, which are used in

decomposition-based algorithms to maintain diversity.

Ø Without considering Pareto dominance relationship, which is inefficiency for MaOPs

Ø Using clustering method to classify the population into a number of clusters, which is expected to solve the difficulty of balancing convergence and diversity in high dimensional objective space.

A clustering-based Many-objective evolutionary algorithms (MaOEA/C)

Page 32: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 32

iq iaiq iaia

MaOEA/C

Two clustering methods are considered:Ø Partitional clustering method (PCM)

Ø Hierarchical clustering method (HCM)

Ø The adopted two-step clustering strategy (PCM followed HCM) aims to efficiently classify the union of parent and offspring populations into N clusters, requiring a computational cost similar to that of most state-of-the-art MaOEAs

Page 33: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 33

MaOEA/C

Partitional clustering method (PCM)• divide a population S into m clusters• m axes serve as the clustering centers, m is the number of objectives• angle value between solutions and axes as the similarity metric

Run Algorithm 1 to divide the solutions set S into m clusters and each cluster is also a solutions set

Page 34: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 34

MaOEA/CHierarchical clustering method (HCM)• divide a population S into k clusters, k is less than the size of S

• initialize each solution in S as the center, and each solution in S is regarded as a cluster

• Finally, only k clusters are remain

Page 35: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 35

MaOEA/C

Hierarchical clustering method (HCM)• angle value between two different centroids as the similarity metric• each time, combine two clusters (with the minimum

angle value between their centroids) into a new cluster , and update its centroid by

• all the computation of angle values and update of centroids are run in the normalized objective space.

Page 36: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 36

MaOEA/C

Pseudo-code of environmentalselection inMaOEA/C

Page 37: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 37

MaOEA/C

Pseudo-code of MaOEA/C

Page 38: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 38

A simple example with bi-objective optimization to illustrate the environmental selection process, using (a) PCM and then followed (b) HCM.

MaOEA/CTwo-step clustering strategy (PCM followed HCM)

In each final cluster, a solution with the best convergence performance is preserved.

Page 39: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 39

MaOEA/C

SimulationResults onMaFs

Page 40: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 40

MaOEA/C

SimulationResults onWFGs

Page 41: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 41

MaOEA/C(a) Average performance rank on 5-, 8-, 10-, 13- and 15-objectives

(b) Average performance rank on MaF1-7 and WFG1-9 problems

Page 42: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 42

MaOEA/C

2.094

3.8884.125

4.381 4.400 4.493 4.619

MaOEA/C EFR-RR Ɵ-DEA VaEA NSGA-III MOEA/D-DU SRA

Average ranking of Friedman’s test for the compared MaOEAs

All considered problems (MaF, WFG) with different objectives (5, 8, 10, 13, 15) are included

[4] Q. Lin,S. Liu, K. Wong, M. Gong, C. A. Coello Coello, J. Chen, J. Zhang, “A Clustering-based Evolutionary Algorithm for Many-objective Optimization Problems,” IEEE Transactions on Evolutionary Computation, in press, 2018

Page 43: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

2019-4-16 43

The average running times of the selected seven MaOEAs on MaF1-MaF7 (M1-M7) and WFG1-WFG9 (W1-W9) with 10 objectives

MaOEA/C

Page 44: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Outline

2019-4-16 44

1 Basic Concepts

Petri网研究现状2 Three Typical MOEAs

3 An Ensemble MOEA Framework

4 A Clustering-based MaOEA

5 Conclusions & Future Work

Page 45: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Conclusions

Ø Three typical MOEAs have their own shortcomings for MOPs and MaOPs

Ø An ensemble framework of combining the advantages of three typical MOEAs can effectively handle MOPs

Ø Embedding clustering methods into evolutionary algorithm is an effective way for MaOPs

2019-4-16 45

Page 46: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Future Work

Ø An effective ensemble framework for MaOPs

Ø Using clustering method to self-guide weight vectors in decomposition-based MOEAs for improving diversity

Ø Effective variation operators to produce new solutions of MaOPs

Ø Embedding other machine learning methods into evolutionary algorithm for solving MaOPs

2019-4-16 46

Page 47: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

个人主页: http://csse.szu.edu.cn/en/people1bbd.html?30298[1] Qiuzhen Lin, Songbai Liu, et al., A Clustering-based Evolutionary Algorithm for Many-objective Optimization Problems, IEEE Transactions on Evolutionary Computation, in press, online: DOI: 10.1109/TEVC.2018.2866927[2] Qiuzhen Lin, Songbai Liu, et al., Particle Swarm Optimization with A Balanceable Fitness Estimation for Many-objective Optimization Problems, IEEE Transactions on Evolutionary Computation, 2018, 22(1), 32-46.[3] Qiuzhen Lin, Jianyong Chen*, et al., A Hybrid Evolutionary Immune Algorithm for Multiobjective Optimization Problems, IEEE Transactions on Evolutionary Computation, Oct. 2016, 20(5), 711-729.[4]Wenjun Wang, Shaoqiang Yang, Qiuzhen Lin*, et al., An Effective Ensemble Framework for Multiobjective Optimization, IEEE Transactions on Evolutionary Computation, in press, online: DOI: 10.1109/TEVC.2018.2879078.[5] Qiuzhen Lin, Genmiao Jin, et al., A Diversity-Enhanced Resource Allocation Strategy for Decomposition-based Multiobjective Evolutionary Algorithm, IEEE Transactions on Cybernetics, 2018, 48(8), 2388-2401.[6] Qingling Zhu(研究生), Qiuzhen Lin *, et al., An External Archive-Guided Multi-objective Particle Swarm Optimization Algorithm, IEEE Transactions on Cybernetics, Sep 2017, 47(9), 2794 – 2808.[7]Fei Chen, Donghong Wang, Qiuzhen Lin *, et al., Towards Dynamic Verifiable Pattern Matching, IEEE Transactions on Big Data, 2018, in press, DOI: 10.1109/TBDATA.2018.2868657.[8]Lijia Ma, Jianqiang Li*, Qiuzhen Lin, et al., Reliable Link Inference for Network Data With Community Structures, IEEE Transactions on Cybernetics, 2018, in press, DOI: 10.1109/TCYB.2018.2860284.

2019-4-16 47

Page 48: Multiobjective and Many-Objective Optimization using …imagelab.stu.edu.cn/upload/files/2019041617315391.pdf · Multiobjective and Many-Objective Optimization using Evolutionary

Thank you!

2019-4-16 48