three generations of automatically designed robots · three generations of automatically designed...
TRANSCRIPT
ThreeGenerationsof AutomaticallyDesignedRobots
JordanB. Pollack,Hod Lipson,GregoryHornby, andPabloFunes
27thAugust2002
DEMO LaboratoryComputerScienceDept.,
BrandeisUniversity,Waltham,MA 02454,USA,[email protected]
Abstract
Thedifficultiesassociatedwith designing,building andcontrollingrobotshave led their developmentto a stasis: applicationsare limited mostly torepetitive taskswith predefinedmoves.Overthelastfew yearswehavebeentrying to addressthischallengethroughanalternativeapproach:Ratherthantrying to controlan existing macine,or createa general-purposerobot,weproposethat both the morphologyand the controller shouldevolve at thesametime. This processcanleadto theautomaticdesignandfabricationofspecialpurposemechanismsandcontrollersthatachieve specificshort-termobjectives.Hereweprovideabrief review of threegenerationsof our recentresearch,underlyingtherobotsshown on thecover of this issue:Automat-ically designedstaticstructures,automaticallydesignedandmanufactureddynamicelectromechanicalsystems,andmodularrobotsautomaticallyde-signedthroughagenerative DNA-lik eencoding.
1
1 Introduction
Thehighcostsassociatedwith designing,building andcontrollingrobotshaveled
their developmentto a stasis[28]. Robotsin industryareonly appliedto simple
andhighly repetitivemanufacturingtasks.Eventhoughsophisticatedteleoperated
machineswith sensorsandactuatorshave found importantapplications(explo-
rationof inaccessibleenvironmentsfor example),they leave very little decision,
if atall, to theon-boardsoftware[29].
Thecentralissueaddressedby ourwork is away to getahigherlevel of com-
plex physicalityundercontrolat lowercost.Weseekmorecontrolledandmoving
mechanicalparts,moresensors,morenonlinearinteractingdegreesof freedom
- without entailingboth thehugefixedcostsof humandesignandprogramming
andthevariablecostsin manufactureandoperation.We suggestthat this canbe
achieved only whenrobot designandconstructionarefully automatic.Consid-
ering a robot asan artificial life form, we achieve automaticdesignby evolving
creatures— body andbrain — throughinteractionwith (simulated)reality, and
transferthedesignsinto (real)reality.
Traditionally, robotsaredesignedonahardwarefirst, softwarelastbasis:Me-
chanicalandelectricalengineersdesigncomplex articulatedbodieswith state-of-
the-artsensors,actuatorsandmultiple degreesof freedom.Thenext taskshould
be simply to “write the software”. But humanshave drasticallyunderestimated
animalbrains:looking into naturewe seeanimalbrainsof very high complexity,
controllingbodieswhich have beenselectedby evolution preciselybecausethey
2
werecontrollableby thosebrains.We believe thatthecostsof writing intelligent
controllersfor arbitrarymechanicaldevicesaresohigh thatdirectengineeringof
robotswill continueto beaneconomicfailure.
In nature,the body andbrain of a creaturearetightly coupled,the fruit of a
long seriesof smallmutualadaptations— like chickenandegg, neitheronewas
designedfirst. Thereis never a situationin which thehardwarehasno software,
or wherea growth or mutation— beyond the adaptive ability of the brain —
survives.Autonomousrobots,like living creatures,requirea highly sophisticated
correspondencebetweenbrain,bodyandenvironment.
Therefore,we have beenworking to co-evolve both the brain andthe body,
simultaneouslyandcontinuously, from asimplecontrollablemechanismto oneof
sufficientcomplexity for aparticularspecializedtask.Althoughwewerenot first
to proposebrain/bodycoevolution [6, 24, 27], we have beenableto put the idea
in practice.
Over thenext decade,weseethreetechnologiesthatarematuringpastthresh-
old to make possiblea new industryof inexpensive automaticallydesignedma-
chines.Oneis the increasingfidelity of advancedmechanicaldesignsimulation,
stimulatedby profits from theCAD softwareindustry[34]. Thesecondis rapid,
one-off prototypingandmanufacture,which is proceedingfrom 3D plasticlayer-
ing to strongercompositeandmetal(sintering)technology[7]. Thethird is most
centralto Artificial Life, our understandingof the dynamicsof coevolutionary
learningin theself-organizationof complex systems[31, 19,1, 8].
3
2 Coevolution
Coevolution, whensuccessful,dynamicallycreatesa seriesof learningenviron-
mentseachslightly morecomplex thanthelast,andaseriesof learnerswhichare
tunedto adaptin thoseenvironments. Sims’ work [33] on body-braincoevolu-
tion andthemorerecentFramstickssimulator[23] demonstratedthat theneural
controllersandsimulatedbodiescouldbecoevolved.Thegoalof our researchin
coevolutionaryroboticsis to replicateandextendresultsfrom virtual simulations
like theseto the reality of computerdesignedand constructedspecial-purpose
machinesthatcanadaptto realenvironments.
We are working on coevolutionary algorithmsto develop control programs
operatingrealisticphysicaldevice simulators,bothcommercial-off-the-shelfand
our own customsimulators,wherewe finish the evolution insidereal embodied
robots. We areultimately interestedin mechanicalstructuresthat have complex
physicalityof moredegreesof freedomthananythingthathaseverbeencontrolled
by humandesignedalgorithms,with lower engineeringcoststhancurrentlypos-
siblebecauseof minimalhumandesigninvolvementin theproduct.
It is not feasiblethatcontrollersfor completestructurescouldbeevolved(in
simulationor otherwise)without first evolving controllersfor simplerconstruc-
tions.Comparedto thetraditionalform of evolutionaryrobotics[9, 5, 26,13,22]
which serially downloadscontrollersinto a given pieceof hardware, it is rela-
tively easyto explore the spaceof body constructionsin simulation. Realistic
simulationis alsocrucial for providing a rich andnonlinearuniverse.However,
4
while simulationcreatestheability to explorethespaceof constructionsfar faster
thanreal-world building andevaluationcould,thereremainstheproblemof trans-
fer to realconstructionsandscalingto thehigh complexities usedfor real-world
designs.
3 Results
We describethreegenerationsof work in our lab towardsfully automatedde-
sign andmanufactureof high-parts-countautonomousrobots. The fundamental
methodis evolution insidesimulation,but in simulationsmoreandmorerealistic
so the resultingblueprintsarenot simply visually believable,as in Sims’ work,
but also buildable, either manuallyor automatically. Our first resultsinvolved
automaticallycreatinghigh part-countstructuresthat could be transferredfrom
simulationto the realworld. In thesecondgenerationwe evolvedautomatically
buildabledynamicmachinesthatarenearlyautonomousin boththeir designand
manufacture,usingRapidPrototypingtechnology. Thethird generationbeginsto
addressscaling,by handlinghigh part-countstructuresthroughmodularity. Even
with thesethreedemonstrations,we feel the work is at a very early stage,with
majorissuesonly beginningto beaddressedsuchastheintegrationof sensorsand
automatingthefeedbackfrom “li ve” interactions.
5
3.1 Generation 1: Legobots
Thefirst steptowardsourvisionof fully evolvedcreatureswasto demonstratethat
evolving morphologyfor therealworld wasindeedpossible.
In orderto evolveboththemorphologyandbehavior of autonomousmechan-
ical devicesthatcanbebuilt, onemusthaveasimulatorthatoperatesundermany
constraints,and a resultantcontroller that is adaptive enoughto cover the gap
betweenthesimulatedandrealworld.
Featuresof a simulatorfor evolving morphologyare:
� Representation—- shouldcoverauniversalspaceof mechanisms.
� Conservative — becausesimulationis never perfect,it shouldpreserve a
margin of safety.
� Efficient— it shouldbequicker to testin simulationthanthroughphysical
productionandtest.
� Buildable— resultsshouldbeconvertiblefrom asimulationto arealobject.
Oneapproachis to custom-build asimulatorfor modularroboticcomponents,and
thenevolveeithercentralizedor distributedcontrollersfor them.
In advanceof a modularsimulatorwith dynamics,we built a simulatorfor
(static)Lego bricks, andusedsimpleevolutionaryalgorithmsto createcomplex
Legostructures,whichwerethenmanuallyconstructed[10, 11,12].
Our modelconsidersthe union betweentwo bricks asa rigid joint between
thecentersof massof eachone,locatedat thecenterof theactualareaof contact
6
Figure1: Photographsof theFAD LegoBridge(Cantilever)andCrane(Triangle).Photographscopyright PabloFunes& JordanPollack,usedby permission.
betweenthem. This joint hasa measurabletorquecapacity:morethana certain
amountof force appliedat a certaindistancefrom the joint will breakthe two
bricksapart.Thefundamentalassumptionof our modelis theidealizationof the
unionof two Legobricksasa rotationaljoint with limited capacity.
The evolutionaryalgorithmreliably builds structuresthat meetfitnessgoals,
exploiting physicalpropertiesimplicit in the simulation. Building the resultsof
theevolutionarysimulation(by hand)demonstratedthepower andpossibilityof
fully automateddesign:thelong bridgeof figure1 shows thatour simplesystem
discoveredthecantilever, while theweight-carryingcraneshowsit discoveredthe
basictriangularsupport.
7
3.2 Generation 2: Genetically Organized Lifelike Electrome-
chanics (GOLEM)
The Lego machines,with computergeneratedblueprints,andmanualconstruc-
tion, demonstratedthat the interactionbetweensimulatedphysicsandevolution
leadsto a primitive form of discovery which canbe transferredinto reality. The
next goalis to addsomemotionto thesemachines,andaddresstheissueof manu-
facture.While Legokits havemotioncomponents,thedesignspaceis verybroad
anddifficult to model,andno robotcanmatchthemanualdexterity of a 10 year
old humanin assembly.
We startedwith a whole new processin which robot morphologywascon-
strainedto bebuildableby a commercialoff theshelfrapidprototypingmachine.
We evolve the bodiesandcontrollersin simulationandwereessentiallyableto
replicatethemautomaticallyinto reality [25].
Theserobotsarecomprisedof only linearactuatorsandsigmoidalcontrolneu-
rons embodiedin an arbitrary thermoplasticbody. The entire configurationis
evolved for a particulartaskandselectedindividualsareprintedpre-assembled
(exceptmotors)using3D solidprinting(rapidprototyping)technology, laterto be
recycled into differentforms. In doingso,we establishfor thefirst time a com-
pletephysicalevolution cycle. In this project,the evolutionarydesignapproach
assumestwo mainprinciples:(a) to minimizeinductivebias,wemuststriveto use
the lowestlevel building blockspossible,and(b) we coevolve the body andthe
control,sothatthatthey stimulateandconstraineachother. Weusearbitrarynet-
8
worksof linearactuatorsandbarsfor themorphology, andarbitrarynetworksof
sigmoidalneuronsfor thecontrol. Evolution is simulatedstartingwith a soupof
disconnectedelementsandcontinuesoverhundredsof generationsof hundredsof
machines,until creaturesthataresufficiently proficientat thegiventaskemerge.
The simulatorusedin this researchis basedon quasi-staticmotion. The basic
principle is that motion is broken down into a seriesof statically-stableframes
solvedindependently. While quasi-staticmotioncannotdescribehigh-momentum
behavior suchasjumping,it canaccuratelyandrapidly simulatelow-momentum
motion.Thiskind of motionis sufficiently rich for thepurposeof theexperiment
and,moreover, it is simpleto inducein reality sinceall real-timecontrol issues
areeliminated. Several evolution runswerecarriedout for the taskof locomo-
tion. Fitnesswasawardedto machinesaccordingto theabsoluteaveragedistance
traveledoveraspecifiedperiodof neuralactivation.Theevolvedrobotsexhibited
variousmethodsof locomotion,includingcrawling, ratchetingandsomeformsof
pedalism(Figure2). Selectedrobotsarethenreplicatedinto reality: their bodies
arefirst fleshedto accommodatemotorsandjoints,andthencopiedinto material
usingrapid prototypingtechnology. Temperature-controlledprint headextrudes
thermoplasticmateriallayerby layer, so that thearbitrarily evolvedmorphology
emergespre-assembledasa solid three-dimensionalstructurewithout tooling or
humanintervention.Motorsarethensnappedin (manually),andtheevolvedneu-
ral network is activated(Figure3). Therobotsthenperformin reality asthey did
in simulation.
9
(a) (b) (c)
(d) (e) (f)
Figure2: (a) A tetrahedralmechanismthat produceshinge-like motion andad-vancesby pushingthe centralbar againstthe floor. (b) Bipedalism:the left andright limbsareadvancedin alternatingthrusts.(c) Movesits two articulatedcom-ponentsto producecrab-like sidewaysmotion. (d) While the uppertwo limbspush,the centralbody is retracted,andvice versa. (e) This simplemechanismusesthe top bar to delicatelyshift balancefrom sideto side,shifting thefrictionpoint to eithersideasit createsoscillatorymotionandadvances.(f) This mech-anismhasanelevatedbody, from which it pushesanactuatordown directly ontothefloor to createratchetingmotion. It hasa few redundantbarsdraggedon thefloor.
10
(a) (b) (c)
(d) (e) (f)
Figure3: (a)Fleshedjoints,(b) replicationprogress,(c) pre-assembledrobot(fig-ure2f), (d,e,f)final robotswith assembledmotors
3.3 Generation 3: Modularity Generative Design (Tinkerbots)
While the GOLEM projectdemonstratedvalidity of our approachto automatic
designandmanufacture,themachineswhich wereproducedareobviously fairly
simplecomparedto the kinds of robotsbuildableby teamsof humanengineers.
In fact mostwork in automaticdesignof engineeringproductsusingtechniques
inspiredby biologicalevolution,[21, 3, 17, 10,4, 25] suffersthesamecriticism.
Our third generationstartsto addressthe issueof whetherevolutionaryauto-
matic designtechniquescanattain the higher level of complexity necessaryfor
practicalengineeringprojects. Sincethe searchspacegrows exponentiallywith
thesizeof theproblem,searchalgorithmsthatusea directencodingfor designs
maynot scaleto largedesigns.An alternative to a directencodingis a generative
11
specification,which is a grammaticalencodingthat specifieshow to constructa
design,[32] and[2]. Similar to a computerprogram,a generative specification
canallow the definition of re-usablesub-proceduresallowing the designsystem
to scaleto morecomplex designsthancanbeachievedwith adirectencoding.
Ideally an automateddesignsystemwould startwith a library of basicparts
andwould iteratively createnew, morecomplex modules,from onesalreadyin its
library. The principle of modularity is well acceptedasa generalcharacteristic
of design,as it typically promotesdecouplingandreducescomplexity [35]. In
contrastto a designin which every componentis unique,a designbuilt with a
library of standardmodulesis more robust and more adaptable,and enhances
field repair.
Our third generationof automaticallydesignedrobotsfocuson modularde-
sign,andusesL-systemsasthegenotypeevolvedby theevolutionaryalgorithm.
L-systemsareagrammaticalrewriting systemintroducedbyLindenmeyerin 1968
to modelthebiologicaldevelopmentof multicellularorganisms.Rulesareapplied
in parallelto all charactersin thestringjust ascell divisionshappenin parallelin
multicellular organisms.Complex objectsarecreatedby successively replacing
partsof a simple objectby using the set of rewriting rules. Using this system
we have evolved3D staticstructures[15], andlocomotingmechanisms[14, 16],
someof whichareshown in figure4, andandtransferredsuccessfullyinto reality,
asseenin figure5 [14]. Thecreatureon thecover is our first 3D machineusing
thesetechniques.
12
Figure4: Examplesof evolved,modularcreatures.
Figure5: Two partsof thelocomotioncycleof a2D,modularlocomotingcreaturein bothsimulationandreality.
13
4 Discussion and Conclusion
Can evolutionary and coevolutionary techniquesbe usedin the designof real
robotsas“Artificial Lifeforms?” In thispaperwehavepresentedthreegenerations
of ourwork, eachof whichaddressesoneor moredimensionsof theproblem.We
haveoverviewedresearchin useof simulationsfor handlinghighpart-countstatic
structuresthat arebuildable,dynamicelectromechanicalsystemswith complex
morphologythatcanbebuilt automatically, andgenerativeencodingsasa means
for scalingto complex structures.
The limitations of the work areclearly apparent:thesemachinesdo not yet
have sensors,andarenot really interactingwith their environments. Feedback
from how robotsperformin therealworld is not automaticallyfed-backinto the
simulations,but requirehumansto refinethe simulationsandconstraintson de-
sign. Finally, thereis the questionof how complex a simulatedsystemcanbe,
beforetheerrorsgeneratedby transferto reality areoverwhelming.
We cannotclaim immediatesolutionto theseproblems.In otherwork, how-
ever, we have demonstratedhow coevolution can lead to complex performance
in domainslikegame-playing[31] anddesignof complex algorithmslikesorting
networks[18] andcellularautomata[20]. In ournext generationsof evolvedcrea-
turesweexpectto seesomesensorintegration,andwehavealreadydemonstrated
robot“cultural” evolution, learningfrom interactingin therealenvironment.[36]
Theissueof whetherornotthiskind of artificial life workwill everbepractical
andscaleableis bestrelatedto the history of computerchess. The theory that
14
machinescouldplayagamelikechesswasfrom the1920’s,thefirst chessplaying
computerwas built in the mid 1950’s, and maderandomlegal moves. While
proponentsof fundingfor thenew field of AI wereover-optimistic,by theendof
the century, nevertheless,with almostunlimited CPU time at its disposal,Deep
Blue wasableto win a tournamentagainstthe leadinghumanplayer- using80
yearold theory[30].
Perhapsthesmalldemonstrationsof automaticdesignwill lead— with con-
tinueddevelopment,andincreasesin computerspeedandsimulationfidelity, cou-
pledto increasesin basictheoryof coevolutionarydynamics— over time, to the
point wherefully automaticdesignis takenfor granted,muchascomputeraided
designis takenfor grantedin manufacturingindustriestoday.
Our currentresearchmovestowardsthe overall goal via multiple interacting
paths,of simulation,theory, building andtestingin the real world, andapplica-
tions. It is a broad,multidisciplinary long-termendeavor, wherewhat we learn
in onepathaidsthe others. We believe sucha broadendeavor is the only way
to ultimately constructcomplex autonomousmachineswhich caneconomically
justify their own existence.
Acknowledgements
This researchwassupportedover theyearsin partby theNationalScienceFoun-
dation(NSF),theofficeof Naval Research(ONR),andby theDefenseAdvanced
ResearchProjectsAgency (DARPA).
15
References
[1] PeterJ. Angeline,Gregory M. Saunders,andJordanB. Pollack. An evolu-tionaryalgorithmthatconstructsrecurrentnetworks. IEEE TransactionsonNeural Networks, 5(1):54–65,1994.
[2] P. Bentley andS. Kumar. Threeways to grow designs:A comparisonofembryogeniesof anevolutionarydesignproblem.In Banzhaf,Daida,Eiben,Garzon,Honavar, Jakiel,andSmith,editors,GeneticandEvolutionaryCom-putationConference, pages35–43,1999.
[3] P. J.Bentley. GenericEvolutionaryDesignof SolidObjectsusinga GeneticAlgorithm. PhDthesis,Divisionof ComputingandControlSystems,Schoolof Engineering,TheUniversityof Huddersfield,1996.
[4] Peter Bentley, editor. Evolutionary Design by Computers. Morgan-Kaufmann,SanFrancisco,1999.
[5] D. Clif f, I Harvey, andP. Husbands.Evolution of visual control systemsfor robot. In M. SrinivisanandS. Venkatesh,editors,FromLiving EyestoSeeingMachines. Oxford,1996.
[6] D. Clif f and J. Noble. Knowledge-basedvision and simple visual ma-chines. PhilosophicalTransactionsof theRoyalSocietyof London: SeriesB, 352:1165–1175,1997.
[7] D. Dimos,S.C.Danforth,andM.J.Cima. Solid freeformandadditivefabri-cation.In J.A. Floro,editor, Growthinstabilitiesanddecompositionduringheteroepitaxy. Elsevier, 1999.
[8] S.Ficici andJ.B. Pollack.Challengesin coevolutionarylearning:Arms-racedynamics,open-endedness,andmediocrestablestates.In Adami,Belew, Ki-tano,andTaylor, editors,Proceedingsof theSixthInternationalConferenceonArtificial Life, 1998.
[9] D. FloreanoandF. Mondada.Evolution of homingnavigationin a realmo-bile robot. IEEE TransactionsonSystems,Man,andCybernetics, 1996.
[10] PabloFunesandJordanB. Pollack.Computerevolutionof buildableobjects.In Phil HusbandsandInmanHarvey, editors,Fourth EuropeanConferenceonArtificial Life, pages358–367,Cambridge,1997.MIT Press.
16
[11] PabloFunesandJordanB. Pollack. Evolutionarybodybuilding: Adaptivephysicaldesignsfor robots.Artificial Life, 4(4):337–357,1998.
[12] PabloFunesandJordanB. Pollack.Computerevolutionof buildableobjects.In PeterBentley, editor, EvolutionaryDesignby Computers, pages387 –403.Morgan-Kaufmann,SanFrancisco,1999.
[13] J. C. Gallagher, R. D. Beer, K. S. Espenschield,andR. D. Quinn. Appli-cationof evolved locomotioncontrollersto a hexapodrobot. RoboticsandAutonomousSystems, 19:95–103,1996.
[14] G. S. Hornby, H. Lipson,andJ. B. Pollack. Evolution of generative designsystemsfor modularphysicalrobots. In IEEE InternationalConferenceonRoboticsandAutomation, pages4146–4151,2001.
[15] G. S. HornbyandJ. B. Pollack. Theadvantagesof generative grammaticalencodingsfor physicaldesign. In Congresson EvolutionaryComputation,pages600–607,2001.
[16] G.S.HornbyandJ.B. Pollack.Body-braincoevolutionusingl-systemsasagenerativeencoding.In GeneticandEvolutionaryComputationConference,pages868–875,2001.
[17] P. Husbands,G. Germy, M. McIlhagga,andR. Ives. Two applicationsofgeneticalgorithmsto componentdesign.In T. Fogarty, editor, EvolutionaryComputing. LNCS1143, pages50–61.Springer-Verlag,1996.
[18] H. Juillé andJ.B. Pollack.Co-evolving intertwinedspirals.In Proceedingsof the Sixth International Conferenceon GeneticAlgorithms, pages351–358,1996.
[19] H. Juillé andJ. B. Pollack. Dynamicsof co-evolutionarylearning. In Pro-ceedingsof theFourth InternationalConferenceon Simulationof AdaptiveBehavior, pages526–534.MIT Press,1996.
[20] H. Juillé andJ. B. Pollack. Coevolving the "ideal trainer": Discovery ofcellularautomatarules. In Koza,editor, ProceedingsThird AnnualGeneticProgrammingConference, July1998.
[21] C. KaneandM. Schoenauer. Geneticoperatorsfor two-dimentionalshapeoptimization. In J.-M. Alliot, E. Lutton, E. Ronald,M. Schoenauer, andD. Snyers,editors,Artificiale Evolution- EA95. Springer-Verlag,1995.
17
[22] Y. Kawauchi, M. Inaba, and T. Fukuda. Genetic evolution and self-organizationof cellular robotic system.JSMEInt. J. SeriesC. (Dynamics,Control, Robotics,Design& Manufacturing), 38(3):501–509,1999.
[23] M. KomosinskiandS. Ulatowski. Framsticks:Towardsa simulationof anature-likeworld,creaturesandevolution. In Jean-DanielNicoudDarioFlo-reanoandFrancescoMondada,editors,Proceedingsof 5th EuropeanCon-ferenceonArtificial Life (ECAL99), volume1674of LectureNotesin Artifi-cial Intelligence, pages261–265.Springer-Verlag,1999.
[24] W. Lee,J. Hallam,andH. Lund. A hybrid gp/gaapproachfor co-evolvingcontrollersandrobot bodiesto achieve fitness-specifiedtasks. In Proceed-ings of IEEE 3rd InternationalConferenceon EvolutionaryComputation,pages384–389.IEEEPress,1996.
[25] Hod Lipson andJordanB. Pollack. Automaticdesignandmanufactureofroboticlifeforms. Nature, 406(6799):974–978,2000.
[26] H. Lund. Evolving robotcontrolsystems.In Alexander, editor, Proceedingsof 1NWGA. Universityof Vaasa,1995.
[27] H. Lund, J Hallam,andW. Lee. Evolving robot morphology. In Proceed-ings of IEEE Fourth International Conferenceon EvolutionaryComputa-tion, pages197–202.IEEEPress,1997.
[28] HansP. Moravec. Riseof the robots. ScientificAmerican, pages124–135,December1999.
[29] J.MorrisonandT. Nguyen. On-boardsoftwarefor themarspathfindermi-crorover. In Proceedingsof the SecondIAA International ConferenceonLow-CostPlanetary Missions. John Hopkins University Applied PhysicsLaboratory, April 1996.
[30] M. Newborn. Kasparov vs. DeepBlue: ComputerChessComesof Age.SpringerVerlag,New York, 1996.
[31] J. B. Pollack and A. D. Blair. Coevolution in the successfullearningofbackgammonstrategy. MachineLearning, 32:225–240,1998.
[32] M. Schoenauer. Shaperepresentationsandevolutionschemes.In L. J.Fogel,P. J.Angeline,andT. Back,editors,Proceedingsof the5th AnnualConfer-enceonEvolutionaryProgramming. MIT Press,1996.
18
[33] K Sims.Evolving3dmorphologyandbehavior bycompetition.In R.BrooksandP. Maes,editors,Proceedings4th Artificial Life Conference, pages28–39.MIT Press,1994.
[34] MarkSincell.Physicsmeetsthehideousbogbeast.Science, 286(5439):398–399,october1999.
[35] NamP. Suh.ThePrinciplesof Design. Oxford UniversityPress,1990.
[36] R. Watson,S. Ficici, and J. B. Pollack. Embodiedevolution: Embody-ing an evolutionary algorithm in a populationof robots. In P. Angeline,Z. Michalewicz, M. Schoenauer, X. Yao, and A. Zalzala, editors, 1999Congresson EvolutionaryComputation, 1999.
19