HAL Id: hal-00297086
https://hal.archives-ouvertes.fr/hal-00297086
Submitted on 15 Jul 2008
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Revisiting the upper bounding process in a safe Branch
and Bound algorithm
Alexandre Goldsztejn, Yahia Lebbah, Claude Michel, Michel Rueher
To cite this version:
Alexandre Goldsztejn, Yahia Lebbah, Claude Michel, Michel Rueher. Revisiting the upper bounding
process in a safe Branch and Bound algorithm. 14th International Conference on Principles and
Practice of Constraint Programming, Sep 2008, Sydney, Australia. pp.598-602. �hal-00297086�
Bran h and Bound algorithm
⋆
AlexandreGoldsztejn1
,YahiaLebbah2,3
,ClaudeMi hel3
,andMi helRueher
3
1
CNRS/UniversitédeNantes2,ruedelaHoussinière,44322Nantes,Fran e alexandre.goldsztejnuniv-nantes .fr
2
Universitéd'OranEs-SeniaB.P.1524EL-M'Naouar,31000Oran,Algeria ylebbahgmail. om
3
UniversitédeNi e-SophiaAntipolis,I3S-CNRS,06903 SophiaAntipolis,Fran e { pjm, rueher}polyte h.uni e.fr
Abstra t. Findingfeasiblepointsforwhi htheproofsu eedsisa rit-i alissueinsafeBran handBoundalgorithmswhi hhandle ontinuous problems.Inthis paper, weintrodu eanewstrategy to omputevery a urateapproximationsof feasible points. This strategytakes advan-tageofthe Newtonmethodfor under- onstrainedsystemsofequations andinequalities.Morepre isely,itexploitstheoptimalsolutionofa lin-earrelaxation ofthe problemto omputee ientlyapromisingupper bound.FirstexperimentsontheCo onutsben hmarksdemonstratethat thisapproa hisveryee tive.
Introdu tion
Optimization problems are a hallenge forCP in nite domains; theyare also a big hallenge for CP on ontinuous domains. The point is that CP solvers are mu h slower than lassi al(non-safe) mathemati al methods on nonlinear onstraint problems as soon as we onsider optimization problems. The te h-niques introdu ed in this paper try to boost onstraints te hniques on these problems andthus, toredu e thegap betweene ientbut unsafe systemslike BARON
1
,andtheslowbut safe onstraintbasedapproa hes.We onsider here theglobaloptimizationproblem
P
tominimizeanobje tivefun tionunder non-linearequalitiesandinequalities,minimize
f
(x)
subje ttog
i
(x) = 0, i ∈ {1, .., k}
h
j
(x) ≤ 0, j ∈ {1, .., m}
(1) withx
∈
x,f
: IR
n
→ IR
,g
i
: IR
n
→ IR
andh
j
: IR
n
→ IR
; Fun tionsf
,g
i
andh
j
arenonlinearand ontinuouslydierentiableon someve torx of intervalsofIR
.For onvenien e,in thesequel,g(x)
(resp.h(x)
)will denotethe ve torofg
i
(x)
(resp.h
j
(x)
)fun tions.⋆
Anextentedversionofthispaperisavailableat:
http://www.i3s.uni e.fr/%7Emh/RR/2008/RR-08.11-A.GOLDSZTEJN.pdf 1
Fun tionBB(INx,
ǫ
;OUTS
,[L, U ]
) %S
:setofprovenfeasiblepoints%
f
x
denotesthesetofpossiblevaluesforf
inx
%
nbStarts
:numberofstartingpointsintherstupper-boundingL←{x}
;(L, U )←(−∞, +∞)
;S←U pperBounding(x
′
, nbStarts)
; whilew([L, U ]) > ǫ
dox
′
←x
′′
su hthatf
x
′′
= min{f
x
′′
: x
′′
∈ L}
;L←L\x
′
;f
x
′
←min(f
x
′
, U)
;x
′
←P rune(x
′
)
;f
x
′
←LowerBound(x
′
)
;S←S ∪ U pperBounding(x
′
,
1)
; ifx
′
6= ∅
then(x
′
1
, x
′
2
)←Split(x
′
)
;L←L ∪ {x
′
1
, x
′
2
}
; ifL = ∅
then(L, U )←(+∞, −∞)
else(L, U )←(min{f
x
′′
: x
′′
∈ L}, min{f
x
′′
: x
′′
∈ S})
endwhileThedi ulties insu h globaloptimization problems ome mainlyfrom the fa tthat manylo alminimizersmayexist butonlyfewofthemareglobal min-imizers [3℄. Moreover, the feasible region may be dis onne ted. Thus, nding feasible pointsis a riti alissue in safeBran hand Boundalgorithms for on-tinuous globaloptimization. Standard strategiesuse lo al sear h te hniques to provideareasonableapproximationofanupperboundandtrytoprovethata feasiblesolutiona tually existswithin the box aroundtheguessed global opti-mum.Pra ti ally,ndingaguessed pointforwhi h theproofsu eeds isoften avery ostlypro ess.
Inthispaper,weintrodu eanewstrategyto omputeverya urate approx-imationsoffeasiblepoints.ThisstrategytakesadvantageoftheNewtonmethod forunder- onstrainedsystemsofequationsandinequalities.Morepre isely,this pro edure exploits the optimal solution of a linear relaxation of the problem to ompute e ientlya promising upperbound.Firstexperiments onthe Co- onutsben hmarksdemonstrate thatthe ombination ofthispro edure witha safeBran handBoundalgorithm drasti allyimprovestheperforman es.
The Bran h and Bounds hema
Thealgorithm(seeAlgorithm1)wedes ribehereisderivedfromthewellknown Bran handBounds hemaintrodu edbyHorstandTuyforndingaglobal min-imizer. Interval analysis te hniques are used to ensure rigorousand safe om-putationswhereas onstraintprogrammingte hniques areused to improvethe redu tionofthefeasiblespa e.
Algorithm1 omputesen losersforminimizersandsafeboundsoftheglobal minimumvaluewithin aninitial boxx .Algorithm1maintainstwolists:alist
L
of boxesto be pro essed anda listS
of provenfeasible boxes.It providesa rigorousen loser[L, U ]
oftheglobaloptimumwithrespe tto atoleran eǫ
.Algorithm1startswith
U pperBounding(x, nbStarts)
whi h omputesaset of feasible boxesby alling alo al sear h withnbStarts
starting pointsand aThebox around the lo al solutionis added to
S
if it is proved to ontain a feasible point. At this stage, if the boxx
′
is empty then, either it doesnot ontain any feasible point or its lower bound
f
x
′
is greater than the urrent upper boundU
. Ifx
′
is notempty, theboxis split alongoneof thevariables 2 oftheproblem.
Inthe main loop, algorithm 1sele tsthe box with thelowest lowerbound of the obje tive fun tion. The
P rune
fun tion applies ltering te hniques to redu ethesizeoftheboxx
′
.Intheframeworkwehaveimplemented,
P rune
just uses a 2B-ltering algorithm [2℄. Then,LowerBound(x
′
)
omputes arigorous lowerboundf
x
′
using a linearprogrammingrelaxation of theinitial problem. A tually, fun tionLowerBound
isbasedon thelinearizationte hniques ofthe Quad-framework[1℄.LowerBound
omputesasafeminimizerf
x
′
thankstothe te hniquesintrodu edbyNeumaieret al.Algorithm1maintains thelowest lowerbound
L
of theremaining boxesL
andthelowestupperboundU
ofprovenfeasibleboxes.Thealgorithmterminates whenthespa ebetweenU
andL
be omessmallerthanthegiventoleran eǫ
.TheUpper-bounding step(seeAlgorithm 2)performsamultistart strategy where a set of
nbStarts
starting points are provided to a lo al optimization solver.Thesolutions omputedbythelo alsolverare thengivento afun tionInf lateAndP rove
whi husesanexisten eproofpro edurebasedontheBorsuk test. However,the proofpro edure mayfail toprovetheexisten eofafeasible point within box xp
. The most ommon sour e of failure is that the guess providedbythelo alsear hliestoofarfrom thefeasibleregion.A newupperbounding strategy
Theupperboundingpro eduredes ribedinthepreviousse tionreliesonalo al sear htoprovideaguessed feasiblepointlyingintheneighborhoodofalo al optima.However,theee tsofoatingpoint omputationontheprovidedlo al optimaare hardto predi t. Asaresult, thelo al optimamightlie outsidethe feasibleregionandtheproofpro eduremightfailtobuildaprovenboxaround thispoint.
Weproposehereanewupperboundingstrategywhi hattemptstotake ad-vantageofthesolutionofalinearouterapproximationoftheproblem.Thelower boundpro essusessu hanapproximationto omputeasafelowerboundof
P
. WhentheLPis solved,asolutionx
LP
isalways omputedand,thus,available forfree.ThissolutionbeinganoptimalsolutionofanouterapproximationofP
, itliesoutsidethefeasibleregion.Thus,x
LP
isnotafeasiblepoint.Nevertheless,x
LP
maybeagood startingpointto onsiderforthefollowingreasons:At ea h iteration, the bran h and bound pro ess splits the domain of the variables.Thesmallertheboxis,thenearest
x
LP
isfromthea tualoptima ofP
.Theproofpro essinatesaboxaroundtheinitialguess.Thispro ess may ompensatetheee t ofthedistan eof
x
LP
from thefeasibleregion. 2Algorithm2Upperboundingbuild fromtheLPoptimalsolution
x
∗
LP
Fun tionUpperBounding(INx,x
∗
LP
,nbStarts
;OUTS
′
) %S
′
:listofprovenfeasibleboxes;
nbStarts
:numberofstartingpoints %x
∗
LP
:theoptimalsolutionoftheLPrelaxationofP(
x)
S
′
← ∅
;x
∗
corr
←
FeasibilityCorre tion(x
∗
LP
); xp
←
InateAndProve(x
∗
corr
,x); if xp
6= ∅
thenS
′
←S
′
∪
xp
returnS
′
However, while
x
LP
onverges to a feasible point, the pro ess might be quite slow. To speed up the upper bounding pro ess, we have introdu ed a light weight,thoughe ient,pro edurewhi h omputeafeasiblepointfromapoint lyingin theneighborhood ofthefeasibleregion.Thispro edure whi his alledF easibilityCorrection
willbedetailedinthenextsubse tion.Algorithm2des ribeshowanupperbound maybebuild fromthesolution ofthelinearproblem usedinthelowerboundingpro edure.
Computing pseudo-feasiblepoints
Thisse tionintrodu esanadaptationoftheNewtonmethodtounder- onstrained systemsofequations andinequalitieswhi hprovidesverya urate approxima-tionsof feasiblepointsat alow omputational ost.Whenthesystemof equa-tions
g(x) = 0
is under- onstrained there is amanifold of solutions.l(x)
, the linear approximation is still valid in this situation, but the linear system of equationsl(x) = 0
isnowunder- onstrained,andhasthereforeananespa eof solutions.Sowehaveto hooseasolutionx
(1)
ofthelinearizedequation
l(x) = 0
among the anespa e of solutions.Asx
(0)
is supposed to bean approximate solutionof
g(x) = 0
,thebest hoi eis ertainlythesolutionofl(x) = 0
whi his the losesttox
(0)
.Thissolution aneasilybe omputedwiththeMoore-Penrose inverse:
x
(1)
= x
(0)
− A
+
g
(x
(0)
)g(x
(0)
)
,whereA
+
g
∈ IR
n×m
istheMoore-Penrose inverseofA
g
∈ IR
m×n
,thesolutionofthelinearized equationwhi hminimizes
||x
(1)
− x
(0)
||
.Applyingpreviousrelationre ursivelyleadstoasequen eof ve -tors whi h onvergesto asolution loseto the initial approximation,provided that thislatterisa urateenough.TheMoore-Penroseinverse anbe omputedinseveralways:asingularvalue de omposition anbeused, orinthe asewhere
A
g
hasfullrowrank(whi h is the aseforA
g
(x
(0)
)
if
x
(0)
is non-singular)theMoore-Penroseinverse an be omputedusing
A
+
g
= A
T
g
(A
g
A
T
g
)
−
1
.
Inequality onstraints are hanged to equalities by introdu ing sla k vari-ables:
h
j
(x) ≤ 0
⇐⇒
h
j
(x) = −s
2
under-InthisSe tion,we ommenttheresultsoftheexperimentswithournewupper bounding strategy ona signi ant set of ben hmarks.Detailled results an be found in the resar h report ISRN I3S/RR-2008-11-FR
3
). All the ben hmarks ome fromthe olle tionofben hmarksof theCo onutsproje t
4
.We have se-le ted
35
ben hmarks where I os su eeds to nd the global minimum while relying on an unsafe lo al sear h. We did ompare our new upper bounding strategywiththefollowingupperbounding strategies:S1: Thisstrategydire tlyusestheguessfromthelo alsear h,i.e.thisstrategy usesasimpliedversionofalgorithm1wheretheproofpro edurehasbeen dropped. As a onsequen e, it doesnot suer from the di ulty to prove theexisten eof a feasiblepoint. However,this strategy is unsafe and may produ ewrongresults.
S2: This strategy attempts to provethe existen e of afeasible point within a boxbuildaroundthelo alsear hguess.Here,allprovidedsolutionsaresafe andthesolvingpro essisrigorous.
S3: Ourupperboundingstrategywheretheupperboundingreliesontheoptimal solutionof the problem linear relaxationto build a box proved to hold a feasible point. A all to the orre tion pro edure attempts to ompensate theee t oftheouterapproximation.
S4: This strategy applies the orre tion pro edure to the output of the lo al sear h(toimprovetheapproximatesolutiongivenbyalo alsear h). S5: This strategy mainly diers from S3 by the fa t that it does not all the
orre tionpro edure
S3, our newupperbounding strategy is the best strategy: 31 ben hmarks are nowsolvedwithinthe30stimeout;moreover,almostallben hmarksaresolved in mu h less time and with a greater amount of proven solutions. This new strategyimprovesdrasti allytheperforman eoftheupperboundingpro edure and ompeteswellwithalo alsear h.
Current work aims at improving and generalizing this framework and its implementation.
Referen es
1. YahiaLebbah,ClaudeMi hel,Mi helRueher,DavidDaney,andJean-Pierre Mer-let. E ientandsafeglobal onstraintsforhandlingnumeri al onstraintsystems. SIAMJournalonNumeri alAnalysis,42(5):20762097, 2004.
2. Olivier Lhomme. Consisten y te hniques for numeri CSPs. In Pro eedings of IJCAI'93,pages232238,Chambéry(Fran e),1993.
3. ArnoldNeumaier.Completesear hin ontinuousglobaloptimizationand onstraint satisfa tion. A taNumeri a,2004.
3
http://www.i3s.uni e.fr/%7Emh/RR/2008/RR-08.11-A.GOLDSZTEJN.pdf 4