evolvability

23
Biologically inspired design of design Ben Bolker Departments of Mathematics & Statistics and Biology, McMaster University 19 December 2010 Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University) Evolutionary computation 19 December 2010 1 / 20

Upload: ben-bolker

Post on 11-May-2015

409 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Evolvability

Biologically inspired design of design

Ben Bolker

Departments of Mathematics & Statistics and Biology, McMaster University

19 December 2010

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 1 / 20

Page 2: Evolvability

1 Introduction

2 Biologically inspired optimization

3 Avenues for exploration/conclusions

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 2 / 20

Page 3: Evolvability

Outline

1 Introduction

2 Biologically inspired optimization

3 Avenues for exploration/conclusions

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 3 / 20

Page 4: Evolvability

Biologically inspired design

Examples:

micro/macro fluid dynamics:kingfisher beaks, robot fish, sharkskin

materials (Velcro, gecko toes)

structural color

Some references:http://www.japanfs.org/en_/newsletter/200503-2.html,

http://www.treehugger.com

http://brainz.org/15-coolest-cases-biomimicry/

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 4 / 20

Page 5: Evolvability

Evolutionary computation

Biologically inspired design of design:i.e., biologically inspired algorithms

Can we learn from evolutionary biology? How?

Generative systems (Genr8, Maya, Rhino . . . )

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 5 / 20

Page 6: Evolvability

Evolutionary computation

Biologically inspired design of design:i.e., biologically inspired algorithms

Can we learn from evolutionary biology? How?

Generative systems (Genr8, Maya, Rhino . . . )

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 5 / 20

Page 7: Evolvability

Problem: bridge design

http://imac.epfl.ch/Team/landolf/Rhode%20et%20al%20EG-ICE%2009.pdf

objective function: cost, performance

parameter space: area of layer and x-cables; outer diameter,diameter-to-thickness ratio of tubular struts; self-stress of layer andx-cables

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 6 / 20

Page 8: Evolvability

Problem: patio design

Caldas (2008) doi:10.1016/j.aei.2007.08.012

objective function: (?)

parameter space: which sides have balconies(24 possibilities, encoded as a bit string): discrete

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 7 / 20

Page 9: Evolvability

Parameter space

http://www.iread.it/lz/hypercube.html

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 8 / 20

Page 10: Evolvability

Outline

1 Introduction

2 Biologically inspired optimization

3 Avenues for exploration/conclusions

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 9 / 20

Page 11: Evolvability

Adaptive landscapes

Wright 1931

(from Johnson 2008)

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 10 / 20

Page 12: Evolvability

“No free lunch” theorem

Across all all possible optimization problems,all optimization algorithms perform equally:none is universally best

. . . a “good” optimization algorithm is only good for some particularproblems

http://en.wikipedia.org/wiki/No_free_lunch_in_search_and_optimization

Ho (2002) http://resolver.scholarsportal.info/resolve/00223239/v115i0003/549_seotntaii.xml

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 11 / 20

Page 13: Evolvability

Consequences of NFL for biologically inspired design

Question

Does biological evolution use good optimization techniques?

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 12 / 20

Page 14: Evolvability

Consequences of NFL for biologically inspired design

Question

///////Does//////////////biological//////////////evolution/////use////////good///////////////////optimization/////////////////techniques?Do evolving systems face the same kinds of problems we do?

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 12 / 20

Page 15: Evolvability

characteristics of objective functions/landscapes

discrete vscontinuous

single vs multiplepeaks

smooth vs jagged

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 13 / 20

Page 16: Evolvability

Selection

Evolution occurs inpopulations

Offspring havedifferentcharacteristics

Best ones survive,the population“climbs the hill”

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 14 / 20

Page 17: Evolvability

mutation

http://en.wikipedia.org/wiki/TMNT

In order to move (and get outof local minima), need tomaintain variation: mutation

too little mutation: slowmovement

too much: constantlylosing fitness

Selection+mutation =“asexualreproduction”

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 15 / 20

Page 18: Evolvability

crossover/recombination

http://www.flickr.com/photos/ajc1/

1103490291/sizes/o/

let individuals “mate”

randomly select somecharacteristics from eachparent

combines features of twodifferent solutions:building blockshypothesis

tradeoff: can also breakup good combinations

modularity is important

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 16 / 20

Page 19: Evolvability

specific algorithms

1

● ●

● ●

●●

50

●●

●●

●●

●●

●●

●● ●

●●●

● ●

●●

● ●●●

● ●●●●●

10

● ●●

●● ●

● ●

● ●

●●

999

●●

●●

●●

●●

●●

●● ●

●●●

● ●

●●

● ●●●

● ●●●●●

genetic algorithmstranslate numeric parametersinto a bit string (e.g. 0011)good for discrete problemsdoes not respect moduleboundaries

differential evolutionnumeric parameters asseparate “genes”different mutation operation— uses direction information

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 17 / 20

Page 20: Evolvability

Outline

1 Introduction

2 Biologically inspired optimization

3 Avenues for exploration/conclusions

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 18 / 20

Page 21: Evolvability

genetic complications (opportunities?)

genetic structure:chromosomes, gene clusters

non-point mutations:deletion, duplication

mating types (♂, ♀)

modifiers: dominance, canalization

genotype-phenotype map:integrating developmental biology(back to generative systems)

Which of these are important foroptimization, and which are accidents?(How and why did they evolve?)

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 19 / 20

Page 22: Evolvability

evolving complexity

closing the loop: development + evolution

can we allow for evolution of complexity (evolving grammars)?

evolution of modularity(adaptive recombination, gene rearrangement)

It’s cool, but is it worth it?

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 20 / 20

Page 23: Evolvability

tradeoffs

general vs. problem-specific solutions (NFL)

performance vs robustness (both in optimization algorithms andsolutions)

programming vs computation time

computation vs “meta-computation”

Ben Bolker (Departments of Mathematics & Statistics and Biology, McMaster University)Evolutionary computation 19 December 2010 21 / 20