25
1 An Evolutionary Algorithm with Species- specific Explosion for Multimodal Optimization GECCO 2009 Ka-Chun Wong, Kwong-Sak Leung, Man-Hon Wong Department of Computer Science & Engineering The Chinese University of Hong Kong, Hong Kong {kcwong, ksleung, mhwong}@cse.cuhk.edu.hk

An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

Embed Size (px)

DESCRIPTION

An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization. GECCO 2009 Ka-Chun Wong, Kwong-Sak Leung, Man-Hon Wong Department of Computer Science & Engineering The Chinese University of Hong Kong, Hong Kong {kcwong, ksleung, mhwong}@cse.cuhk.edu.hk. Who am I ?. - PowerPoint PPT Presentation

Citation preview

Page 1: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

1

An Evolutionary Algorithm with Species-specific Explosion for

Multimodal Optimization

GECCO 2009

Ka-Chun Wong, Kwong-Sak Leung, Man-Hon WongDepartment of Computer Science & Engineering

The Chinese University of Hong Kong, Hong Kong{kcwong, ksleung, mhwong}@cse.cuhk.edu.hk

Page 2: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

2

Who am I ? Name: Wong Ka Chun (黃家駿 ) Position: M.Phil. Student (yr1) in The Chinese University of Hong Kong (CUH

K)

Biography• Ricky received his B.Eng. in Computer Engineering from United College, the Chines

e University of Hong Kong in July 2008. Since August 2008, he has been a M.Phil. student at the Department of Computer Science and Engineering, the Chinese University of Hong Kong, under the co-supervision of Professor LEUNG Kwong-Sak and Professor WONG Man-Hon.With exposures to different aspects (academic, industrial, spiritual and social), he hope that he can strike a balance between them and contribute to the society by researches.

Research Interests• Evolutionary Algorithms• Bioinformatics• Geographical Information System

My Website:• http://www.cse.cuhk.edu.hk/~kcwong/

Page 3: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

3

Motivation

Given an optimization problem, traditional optimization algorithms can be applied to obtain the global optimum.

However, in the real world, we are often interested in not only a single global optimum, but also other possible global and local optima.

Page 4: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

4

Problem Definition

Given a function, an algorithm should work out all optimal points in a single run.

Six-hump Camel Back Function (http://www.it.lut.fi)

Page 5: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

5

Previous works

AEGA (Leung et al. 2003) SCGA (Li et al. 2002) Crowding (Kenneth De Jong 1975) Fitness Sharing (Goldberg et al. 1989) CrowdingDE (R. Thomsen 2004) SDE (Xiaodong Li 2005) Repeated iterations (Beasley et al. 1993) ……

Page 6: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

6

Species-Conserving Genetic Algorithm (SCGA)

Select species seeds• Each species is a subset of population. • The fittest individual within a species is chosen

as the species seed. The region around a species seed forms its corresponding species region.

Page 7: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

7

Species-Conserving Genetic Algorithm

Main idea• Species seeds can bypass the genetic operations

Page 8: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

8

Observation:• For each area (species), only one elitist

(species seed) is saved.• Is it significant enough to help each species to

converge to its respective optimum?

Species-Conserving Genetic Algorithm

Page 9: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

9

Species-Conserving Genetic Algorithm

Not enough individuals to coverge in this species!

Page 10: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

10

Proposed Method

Evolutionary Algorithm with Species-specific Explosion (EASE)• An extension of SCGA

Page 11: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

11

Proposed Method (EASE) Main idea:

• To exploit species seeds for convergences

Page 12: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

12

Proposed Method (EASE)

Species-specific explosion• The local operation in which we create multiple

copies for each species seed and mutate them.

http://www.japanairsoftguns.com/store/images/products/mkkt2shotgun.jpg http://pixal.us/img/up/index/1271575147.jpg

http://www.japanairsoftguns.com/store/images/products/mkkt2shotgun.jpg

http://en.wikipedia.org/wiki/File:Shotgun-shot-sequence-1g.jpg

Page 13: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

13

Proposed Method (EASE)

Species-specific explosion

Page 14: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

14

Proposed Method (EASE)

Species-specific explosion• Two parameters

Number of mutated copies The step-size for mutation

Page 15: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

15

Proposed Method (EASE)

To determine number of mutated copies• No. of mutated copies = pop_size*K*weight

where K is a constant (similar to generation gap)

No. of mutated copies = No. of yellow individuals

Page 16: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

16

Proposed Method (EASE)

To determine the step-size for mutation• Use the last known improvement step size.

step-size for mutation = the average distance from the species seed (blue individual)

Page 17: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

17

Experiments All algorithms were run up to maximum 50000 fitness function evaluati

ons. The performance measurements are obtained by taking the average and standard deviation of 30 runs.

The parameter settings and results of some algorithms* are taken from:• R. I. Lung, C. Chira, and D. Dumitrescu. An agent-based collaborative evol

utionary model for multimodal optimization. In GECCO ’08: Proceedings of the 2008 GECCO conference companion on Genetic and evolutionary computation, pages 1969–1976, New York, NY, USA, 2008. ACM.

• R. I. Lung and D. Dumitrescu. A new evolutionary model for detecting multiple optima. In GECCO ’07: Proceedings of the 9th annual conference on Genetic and evolutionary computation, pages 1296–1303, New York, NY, USA, 2007. ACM.

* The algorithms include: RACE, RO, AEGA, CRDE

Page 18: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

18

Experiments

Performance measurement

Page 19: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

19

Experiments

Page 20: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

20

Experiments

Page 21: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

21

Conclusion

The experimental results reveal that EASE can provide better performance than the others over the benchmark functions. However, it should not be taken to mean that the proposed method (EASE) is “better” than other evolutionary algorithms for multimodal optimization. Such a conclusion is oversimplified.

It shows that EASE improves SCGA for locating optima (global and local), in terms of peak ratio and accuracy without the addition of manual parameters

Page 22: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

22

Future works

1. Better methods to adaptively select1. Number of mutated copies

2. The step-size for mutation

2. The concept Species-specific Explosion will be investigated to improve other evolutionary algorithms.

3. Further experiments will be conducted on high dimensional problems.

Page 23: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

23

Thank you!謝謝 !

Page 24: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

24

Q&A

Page 25: An Evolutionary Algorithm with Species-specific Explosion for Multimodal Optimization

25

Roaming Agent-Based Collaborative Evolutionary Model (RACE)

ROAMING OPTIMIZATION(RO)