Enclosure Methods for Multivariate Di erentiable Functions and Application to Global Optimization

Fred eric Messine

(Lab. LIA, Departement dInformatiqu e de lUniversite de Pau messine@univ-pau.fr)

Jean-Louis Lagouanelle
(Lab. LIMA, Institut de Recherche en Informatique de Toulouse lagouane@irit.fr)

Abstract: The e ciency of global optimization methods in connection with interval arithmetic is no more to be demonstrated. They allow to determine the global optimum and the corresponding optimizers, with certainty and arbitrary accuracy. One of the main features of these algorithms is to deliver a function enclosure dened on a box (right parallelepiped). The studied method provides a lower bound (or upper bound) of a function in that box throughout two di erent strategies. As we shall see, these algorithms associated with various Branch and Bound methods lead to accelerated convergence and permit to avoid the cluster problem. K e y Words: Globaloptimization,Intervalarithmetic,Multivariatefunctions,Branc h and Bound algorithm, Taylors expansion, Polyhedral cone.

Introduction

I n this paper the methods which are proposed, concern the minimization problem i n the context of Branch and Bound algorithms and interval arithmetic [6], [10]. We use a rst order Taylor s expansion and the mean value theorem t o build a ne underestimations of the function over the box. The intersection of these underestimations is easily computed, then a lower bound of the function on the box follows. This lower bound is often better than those which are issued from di erent inclusion functions, i n connection with interval arithmetic [9]. First, we shall deal with the building of the ane underestimations and the principle of the algorithm t o get the lower bound. Then, its eciency will be shown i n section 5, when used within Branch and Bound algorithm, on simple examples of unconstrained minimization of multivariate functions. Usually, the minimization problem is described by: Minimize f ( x ) , subject t o x X , with X IRn . (1)

f is a di erentiable n-variables function and X is the right parallelepiped i n IRn , U dened by X = { x = ( x1 , , xn ) ; xL i xi xi }, i {1 , 2, ,n } . The principle used here, including the underlying mean value theorem, has already been studied i n the case of univariate polynomial function by Jaumard, Hansen, and Xiong [5], by Alefeld [1], and by Visweswaran and Floudas [11]. Mladineo gives, for functions that satisfy a Lipchitz condition, a method [8] based on the building of cones with spherical bases, where a mixed linear nonlinear system must be solved.

I n our method, we get polyhedral cones and we want t osolve very simple linear systems t o keep the eciency of the algorithms. The unique solution of each linear system corresponds t o the vertex of a polyhedral cone that is meant for a minimizer of f on a box X . This is reached by two suitable choices of vertices of the hypercube as explained i n sections 3. Convergence and complexity properties are given i n section 4. Numerical experiments and applications t o global optimization follow i n section 5. The following section gives the basic tools for the construction of the polyhedral cones.

Underestimating a ne functions

According t o Taylor s expansion of a function f , we get some inclusion function of f i n the box X [9]. This inclusion function is called the Taylor form:

( x,y ) X 2 , f ( y ) f ( x ) + ( X x )T g ( X ) ,
Where g ( X ) is an interval vector of the gradient inclusion of f ; g will be automatically calculated by automatic di erentiation [3], [6]. U L U Let X i = [ xL i , xi ] and G i ( X ) = [gi ( X ) , g i ( X ) ] for i = 1 , 2 , , n . f So, for every x X , ( x ) Gi ( X ) . xi We assume, for the following, that giL ( X ) . giU ( X ) < 0. Let s = ( s1 , , sn ) be the coordinates of a vertex S of the box X and
S S gS ( X ) = (g1 ( X ) , , gn ( X ) )T

where
S gk (X) = L sk xU k gL ( X ) + sk xk gU ( X ) , k { 1 , 2 , , n } U k L k xL xU k xk k xk

(2)

Then, from the following trivial inequalities:

L L L U ( yi xL i )g i ( X ) ( yi xi ) ( yi xi )gi ( X ) U U U L and ( yi xU i )gi ( X ) ( yi xi ) ( yi xi )gi ( X )

for all i = 1 , 2 , , n and any given G i ( X ) , we can deduce for every vertex S of the box X , an ane underestimation of a function f over X , that is f ( y ) f ( s ) + ( y s )T gS ( X ) , y X (3)

So we can obtain 2n ane underestimations of the function f for the box X . Each one of these hyperplanes is a maximum supporting hyperplane a t the vertex S, for the box X .

Global minimum enclosure methods

The rst aim of the following algorithm is t o get a lowerbound of a function f on a box X . This will be achieved by the intersection of n + 1 hyperplanes chosen amongst the 2n hyperplanes that we can build with the properties given i n section 2; but that does not give automatically a lower bound of f on X ; except i n the case of an univariate function [5], [11]. For k = 1 , , 2 n let k be a maximum supporting hyperplane, u k its ane + representation and E k the restriction t o the box X of the associated half-space dened by + Ek = { ( x , z ) IRn I R ; z u k ( x ) , x X } classically we have the Propert y 1 I f
k K + Ek is a polyhedral cone with vertex C which contains the

graph of the function f then for | K | = n + 1 , { C } =

k K

k exists and is unique.

Remark. Thechoiceof |K| = n +1isminimal,itispossibletoset |K| > n + 1 w i t h an analogous convexity property for a polyhedrical set but then the computation of the lower bound becomes too expensive. One way t o obtain existence and uniqueness of such a point C is given by the following choice of the hyperplanes k . First of all the n + 1 vertices Sk selected are not contained i n the same hyperplane of IRn . Then each ane function u k satise s the relations u k ( sk ) = f ( s k ) uk ( x ) f ( x ) , x X u k ( sk ) u k ( x ) , x X (4) (5) (6)

Conditions (4) and (5) follow directly from the fact that k is a maximum supporting hyperplane. Condition (6) supposes, here, that according t o section 2, giL ( X ) giU ( X ) < 0 for all i ; this is obtained by application of a monotonicity testt o f on X . Let ( xc , zc ) IRn IR the coordinates of the point C and u k ( x ) = f ( sk ) + ( x sk ) T gSk ( X ) , we introduce now the Deni tion1 . A set of n + 1 vertices of X is said admissible i f the vertex C of the polyhedral cone dened by n + 1 hyperplanes, built as above, satise s zc f ( x ) , x X and x c X . For commodity the simplex, convex hull of the vertices S0 S1 Sn is said admissible. Such a point C gives a lower bound zc of f and x c is destined t o be a current minimizer for problem (1). Let us see on a short example that any arbitrary set of n + 1 vertices of a box X i n IRn may not be admissible.

Example1. f : X IR2 IR ( x1 , x2 ) f ( x1 , x2 ) = x 2 1 2 x1 x 2 + 3 x1 5 x2 X = [5,5][15,10] then G ( X ) = ([ 27 , 43] , [ 15 , 5])T Let s1 = ( 5 , 1 5 )T , s2 = (5 , 15)T , s3 = (5 , 10)T . For these vertices and u 1 , u2 , u3 the corresponding hyperplanes u 1 ( x1 , x2 ) = f ( 5 , 1 5 ) 2 7 ( x1 + 5 ) 1 5 ( x 2 + 1 5 ) u 2 ( x1 , x2 ) = f ( 5 , 1 5 ) + 4 3 ( x 1 5 ) 1 5 ( x 2 + 1 5 ) u 3 ( x1 , x2 ) = f ( 5 , 1 0 ) 2 7 ( x 1 + 5 ) + 5 ( x2 1 0 ) we get ( xc , zc ) = ( 3 . 57 , 15 , 103 . 6) but f (5 , 10 ) = 110, hence zc is not a lower bound; { s1 , s2 , s3 } is not admissible. I n sections 3.1 and 3.2 we deal with two kinds of admissible sets of vertices; i n section 3.2 we show that such sets always exist and can be easily identied . 3.1 Admissible right simplexes

The rst idea t o get straightforward formulae from linear systems is t o consider the n + 1 vertices of a right simplex generated by a reference vertexS 0 and by the n adjacent vertices of a box X . So doing the linear system is easily solved because from the vertex S0 t o one of the others Sk , only one variable is modied.

S3 S2

S0

S1

Figure 1: Right Simplex The following result gives a su cient condition for a right simplex t o be admissible.

Theore m2. Let S0 be a vertex of a box X i n IRn ,let S 1 , S2 , , Sn be the vertices adjacent to S0 . I f gSk ( X ) is dened by (2) for all k . Then the corresponding right simplex T0 is admissible i f and only i f
n

giS0 ( X ) ( X ) giS0 ( X )

gSi i=1 i

(7)

Sk U U where giSk ( X ) = giL ( X ) i f ( sk ) i = x L i and gi ( X ) = gi ( X ) i f ( sk ) i = x i . k k T Proof. Let sk = ( sk be the coordinates of the vertex Sk for all k 1 , s 2 , , sn ) and u k the ane function

u k ( x ) = f ( sk ) + ( x sk ) T gSk ( X ) First, note that Sk di ers from S0 only by one coordinate, supposed t o be x k . Consider the auxiliary problem Minimize z + subject t o ( x , z ) E k , k K xX (8)

I f x is a solution of this linear program, then ( x , z ) satise s property 1; z is a lower bound of f on the box X and x X . Standard form of (8) is Minimize z subject t o z = ek + f ( sk ) + ( x sk ) T gSk ( X ) , + 0 ek , k = 0 , 1 , 2 , , n ( x , z ) E k , k K xX

(9)

ek is a slack variable. The optimal solution of (9) is reached a t the unique extremal point C , vertex of the polyhedral cone: C =
n

k
k=0

I f x 1 , x2 , , xn are the basis variables and ek = 0 , k = 0 , 1 , , n , x = T ( x 1 , x2 , , xn ) is solution of the linear system z = f ( sk ) + ( x sk ) T gSk ( X ) , k = 0 , 1 , 2 , , n Moreover Sk is adjacent t o S0 for all k and gSk is built following (2). Then by subtraction of equation (k ) from equation (0) one gets directly x k = f ( s0 ) f ( s k ) +
Sk 0 S0 sk k gk ( X ) sk gk ( X ) Sk S0 gk ( X ) gk (X)

Sk S0 gk ( X ) gk (X) k = 1,2,,n

(10)

x X that is proved using (3) and the eventually optimal value z for the problem (8) is z = f ( s0 ) + ( x s ) T gS0 ( X ) (11)

B u t this solution is optimal i f and only i f the marginal costs are non negative. From the standard form (9) we get
n

z = z +
k=0 n

cmk ek
S0 gk (X)

with cm 0 = 1 +
k=1

S0 Sk gk ( X ) gk (X)

and cm k =

S0 gk (X) Sk gk S0 ( X ) gk (X)

, k = 1,,n

L ( X ) g U ( X ) < 0 induces that cm > 0 for any k = 0 and but the assumption gk k k optimality of ( x , z ) implies that cm 0 > 0. Let w ( Gi ( X ) ) : = giU ( X ) giL ( X ) and < G i ( X ) > : = min {g iU ( X ) , | g iL ( X ) | } be respectively the width and the mignitude of the interval G i ( X ) , for i = 1 , 2 , , n .

The following condition of existence of an admissible right simplex is deduced from the worst case i n the previous inequality (7). Theore m3. For n 3 admissible right simplex exists on the box X i f and only if n < Gi ( X ) > 1 (12) w ( Gi ( X ) )
i=1

Remark. As i t can be seen i n Algorithm 1 when no admissible right simplex can be found on the box, this condition is weakened anda lower bound will be found again, but with a loss of accuracy. For univariate or bivariate functions, inequality (12) is always satised and a right simplex can be found. This leads t o the Algorithm 1 Application of a monotony test (elimination of components). Choose a right simplex satisfying (12); determination of the set K . For n 3 if no such simplex exists then apply k one of these local strategies : relaxation of the gradient; then modify until the condition is satised, - Return the intersection of the polyhedral cone with the unbounded cylinder which has for base the box X : return thek lowe r point so obtained. Compute x , z by (10) and (11). k K Relaxation: one moves some hyperplanes k until the condition (7) is satisProof. i f the optimality condition (7) is satised then a lower bound is directly obtained by computation of { C } = , ed, Intersection is equivalent t o compute directly the rst order Taylor inclusion function [10].

This algorithm runs only for right simplexes; theaim of the next section is t o answer positively the question : for a given box X IRn , is i t always possible t o nd an admissible set of n + 1 vertices? 3.2 Construction of admissible simplexes

L L U U U Let x L = ( xL = ( xU 1 , x2 , , xn ) and x 1 , x2 , , xn ) be the vertices of X with is called the opposite of a given vertex S on the box X extremal coordinates. S when s = x L + xU s . We naturally associate t o the box X , the symmetric directed graph constructed from its vertices and edges. Then the following result shows that, for any given vertex S and the opposite one S on the box X , i t is always possible t o nd a path of length n : SS k 1 S k n 1 S such that the corresponding simplex is admissible.

S 1

.. .. .. .. ... ... .. .. .. .. .. .. .. .. ... .. ... .. .... . .... ... ..... ... ....... ... . . ..

S3 33=S 0

S 2
2

S2

.. ... .. .... .. .... .. .... .. .... ......... ... .... .... .. .... .. .... .. .... .....

. .. .... .. .... .. .. .. .. .... .... ...

....... .... .. ....... . ........... ..

S0 =S 3 00

S1 11

Figure 2: Admissible Simplex

giS ( X ) being dened by (2), we denote Ki = |giS ( X ) | , i = {1,,n} w ( Gi ( X ) )

Theore m4. Let S be an arbitrary vertex of the box X , S its opposite. Let the real numbers K i be sorted i n increasing order so that 1 < K k 1 K k 2 K kn < 0 Then the simplex dened by vertices SS k 1 S k n 1 S k n ( = S ) is admissible when the path from S to S is dened for j = 0 , 1 , 2 , , n 1 by : = si j , i { 1 , 2 , , n } \ kj + 1 si kj kj+1 U sk j + 1 : = x L k j + 1 + xk j + 1 sk j + 1
kj+1 k

(13)

 where S k 0 = S and S k n = S. This path is unique for strict inequalities. Proof. One proceeds analogously t o the previous theorem. The only di erence lies on the fact that i n this case, two consecutive vertices have n 1 identical coordinates. Consequently x optimal solution of the associated linear program is obtained directly by subtraction of two consecutive equations; and then, we get x kj =

f ( sk j 1 ) f ( s k j ) + gk j j ( X ) sk jj gk j gk j j ( X ) gk j for k j = 1 , 2 , , n
Sk Sk j
1

Sk

Sk j

( X ) sk jj 1

(14)

(X)

and z = f ( s ) + ( x s )T gS ( X )

(15)

one must show that x X and that z is a lower bound of f . U x X is straightforward when one writes x L k j x k j x k j , for all k j under
L U the hypothesis gk gk j < 0 and (3). j Otherwise z is the minimal value of auxiliary problem (8) i f and only i f the marginal costs are non-negative because n

z = z + ek 0 + which implies that

ek j 1 ek j (X)
S Sk gk j j
1

Sk j k j = 1 gk j

(X)

gk jk 0 ( X )

0< 1 + and
Sk

gk 1k 0 ( X )
k1 gk 1 ( X ) gk 1k 0 ( X )

gk j 0 ( X ) gk j j ( X ) gk j
Sk j
1

Sk

(X)

0 (X) gk j + 1 j + 1 ( X ) gk j +j 1 ( X ) gk j + 1

Sk

Sk

Sk

for j = 1 , 2 , , n 1 B u t considering (13) and (2) we get gk j and

Sk gk j j Sk j
1

( X ) = gk j

Sk j

( X ) = = gk jk 0 ( X )

(X) =

gk j

Sk j U gk j

L (X) ( X ) gk j L gk j

(X)

(X)

L gk j

(X)+

U (X)g gk kj j U gk j

Sk j L gk j

(X)

(X)

(X)

U gk (X) j

for j = 1 , 2 , , n . Therefore, i n any case, gk j 0 ( X ) gk j j ( X ) gk j

Sk Sk j
1

Sk

= (X)

|gk j 0 ( X ) | w ( Gk j ( X ) )

Sk

because gjL ( X ) < 0 < gjU ( X ) , for j = 1 , 2 , , n . Hence for, any initial vertex S, a t least one admissible simplex, can be found . applying the rule given i n theorem(4) that gives a path from S t o S instead of S, i t is easily seen that one nds the I f the initial vertex is S same path i n reverse order towards S. Then, for a box X , i t exists a t least 2n 1 admissible simplexes and as much lower bounds for f . We do not study here the best choice amongst all these possibilities, a choice that must take into account the values of f and its derivative. For illustration, we apply this method t o example 1 43 s = (5 , 15)T is the initial vertex, gS = (43 , 15)T and K 1 = , K2 = 70 15 which gives 20 1 < K2 < K1 < 0 then an admissible path from s = (5 , 15)T t o s = ( 5 , 10)T is obtained by modifying rst the coordinate x 2 which leads t o the vertex (5 , 10) and secondly the coordinate x 1 which leads t o ( 5 , 10) the terminal vertex. 8 2211 One gets with the corresponding hyperplanes x = ( , 5 ) and z = 7 7 315.9. Now we see why the right simplex { s1 , s2 , s3 } was not admissible i n this example; generally for a bivariate function on a box X , two simplexes are admissible; which, i n this case, coincides with the right ones. To get a lower bound of a function f from theorem (4), we apply Algorithm 2 Choose a vertex S; determine gS ( X ) . Application of monotony test (elimination of components). Computation of Kk =
U( gk S(X )| |gk , k = {1,,n} L(X) X ) gk

Classify the associated numbers K k i n increasing order. Determination of an admissible simplex following theorem (4) Compute { x , z } with (14) and (15). Remark. Several very ecient improvements may be used t o get tighter lower bounds: Slope matrices will advantageously take place of derivatives when their computation is possible or narrower intervals for the components of g may also result from Taylor expansion of E. Hansen [3], where some interval arguments are replaced by real quantities. I f an upper bound is required instead of a lower bound, Algorithm 2 is modied as follows For a given vertex S of the box X , let H i = | Ki |, for i = 1 , 2 , , n . giS ( X ) =
L si xU i gU ( X ) + si xi giL ( X ) U i L U x i xi x i xL i

(16)

The real numbers H i are classied i n increasing order 0 < H in H i n 1 H i 1 < 1 (17)

is found by modifying the coordinates of Then an admissible path from S t o S S i n the order x i 1 , xi 2 , , xi n . Unfortunately, the paths for a lower bound and an upper bound of a function f on a box X are distinct excepted i n extreme cases for example i f giL ( X ) = C 1 and giU ( X ) = C 2 for all i . When inequalities are strict i n (17), the two paths are opposite which means are respectively the initial vertices for lower and upper bounds that S and S then Sk i is the opposite vertex of Sk i i.e. s k i = x L + xU sk i for all i, see gure (2).

Optimization algorithm using these enclosure processes

The methods previously described are integrated, by the mean of Algorithm 1 and Algorithm 2 i n interval Branch and Bound algorithms [10] t o solve the global minimization problem (1). This is the topic of Algorithm 3 Initial enclosure of the minimum on X , say [ F L , F U ] Insert ( X , F L ) i n a list L f + While the minimizers are not found with a given tolerance Do - Extract the rst element from L , write i t as follows: ( Y ,y ) - Divide Y into two parts (along its larger edge): we get V ( 1 ) and V ( 2 ) . - For i = 1 , 2 - Do - Enclosure the global minimum on V ( i ) = [ F L , F U ] - if F L > f then - Go t o the next step - else - if F U < f then . f F U . Remove all the elements of L which lower bound is greater than the new f. - end i f - Insert (V ( i ) , F L ) i n L following the increasing order of the F L . - end i f - end For End While Return f and L . Remark. The upper bound F U may also be progressively computed i n the algorithm We have three possibilities t o enclose the global minimum:

1. right simplex and, i f necessary, intersection with unbounded cylinder which has for base thebox X . 2. right simplex and, i f necessary, gradient relaxation. 3. research of an admissible simplex with two opposite vertices of X . When the size of a box X is suciently small, i t is not necessary t o compute for each sub-box of X thegradient according t o inclusion monotonicity but this induces a slower convergence. For every admissible simplex dened by the vertices Sk 0 , Sk 1 , , Sk n of X , the following inequalities hold:
j = 1 , 2 , ,n

min

{ f ( Sk j ) } f m i n z

(18)

with f m i n = inf x X 4.1 Complexity and convergence

Each underestimation of f requires an interval inclusion function of grad f n + 1 point evaluations of the function f and nally the computation of { x , z } requires O (3 n ) elementary operations for a given box X . This estimate is relative t o a single box; but considering, the splitting of a box into adjacent sub-boxes, , which is done i n global optimization problems, the number of evaluations of f may be perceptibly reduced. Indeed i f n vertices belong t o the interface of two sub-boxes, then one more vertex is necessary on each sub-box. Moreover changing a single vertex does not a ect all thecoordinates of x and brings only a partial contribution i n the computation of z . Let zL and zU be respectively the computed lower andupper bounds. Let w ( X ) and w ( G ) be the width vectors, w ( X ) = x U xL and w ( G ) = ( w ( G1 ( X ) ) , w ( G2 ( X ) ) , , w ( Gn ( X ) ) ) T and
S S S |gS ( X ) | = (|g 1 ( X ) | , | g2 ( X ) | , , | gn ( X ) | )T

where gS ( X ) is dened by (2), then properties of convergence of ane estimates run from Theore m5. For a di erentiable function f : X IRn IR; i f an admissible and the opposite one { S k } j = 0 , 1 , ,n to get set { Sk j } j = 0 , 1 ,,n is used to get zL j for z on the box X , then, assuming gU ( X ) g L ( X ) < 0 , k, the following zU U k k inequalities are valid:
f |zL m i n|
j = 0 , 1 , ,n

min

( w ( X ) , | g Sk j ( X ) | )IR n

(19) (20)

z ( w ( X ) , w ( G )) n zU IR L

And the excess width of a ne estimation is at least of order 2 for a regular function.

Proof. The proof is quite straightforward drawn from the arguments : For all j ,
T Sk j |zL fm i n | |zL f ( s k j ) | = ( x (X) L sk j ) g

. and on equivalent inequality for zU For any x X ,

f ( x ) = giL ( X ) + ( 1 ) giU ( X ) , ]0,1[ xi Finally, E W excess width of ane estimate for a function f is given by
E W = ( zU zL ) ( fm ax fm i n ) = zU fm ax ( fm i n zL ) ( w ( X ) , w ( G )) IR n

which induces a convergence of order 2 as soon as, for example, the rst derivative of f satise s a Lipschitz condition.

Numerical experiments

First we search the global minimum of polynomial functions of 2 , 3, or 4 variables with di erent degrees and of functions involving trigonometric functions.
2 f 1 ( x ) = 1 + ( x2 1 + 2 ) x2 + x1 x 2 , with X = [1 , 2] [ 10 , 10] 1 6 4 2 f 2 ( x ) = 2 x2 x , with X = [ 2 , 4]2 1 1 . 0 5 x 1 + x2 x1 x 2 + 6 2 f 3 ( x ) = ( x1 2 x2 7 )2 + ( 2 x1 + x2 5 ) , with X = [ 2 . 5 , 3 . 5] [ 1 . 5 , 4 . 5] 2 f 4 ( x ) = [ 1 + ( x1 + x2 + 1 )2 ( 1 9 1 4 x 1 + 3 x2 1 1 4 x2 + 6 x1 x 2 + 3 x2 )] 2 [ 3 0 + ( 2 x 1 3 x2 ) 2 ( 1 8 3 2 x 1 + 1 2 x2 1 + 4 8 x2 3 6 x1 x 2 + 2 7 x2 )]

with X = [ 2 , 2]2
3 f 5 ( x ) = ( x1 1 ) ( x 1 + 2 ) ( x 2 + 1 ) ( x 2 2 ) x2 3 , with X = [ 2 , 2] 2 2 2 f 6 ( x ) = 4 x2 1 2 x1 x 2 + 4 x2 2 x2 x 3 + 4 x3 2 x3 x 4 + 4 x4 + 2 x1 x2 + 3 x3 + 5 x4 , with X = [ 1 , 3 ] [ 1 0 , 1 0 ] [ 1 , 4 ] [ 1 , 5 ] 2 f 7( x ) = x 2 1 + x2 cos18 x1 + x1 sin18 x2 + x3 cos x3 + x1 x 2 x 3

with X = [1,500]3
Pb N NE 1550 T1 114 T 1 + M 27 RS 64 RS+M 23 f1 f2 T(s) C N T(s) C N 16,80 286 649 7,73 150 218 2,29 9 158 3,76 21 174 0,32 5 18 0,27 6 23 0,65 8 87 1,33 25 55 0,29 5 18 0,29 6 19 f3 T(s) 1,08 1,64 0,23 0,51 0,16 f2 f4 C N T(s) C N T(s) 36 - - 456 4,02 5 5373 356,73 42 331 8,16 3 4769 129,4 38 25 0,53 3 3379 84,83 24 131 1,76 3 3318 83,5 24 22 0,50 C 83 26 8 32 9

. N: Number of iterations, . T(s): CPU time i n seconds,

. . . . .

C: Number of elements enclosing the optimizers i n the list L , -: the algorithm does not give any solution after 15 minutes, NE: Classical Ichida Fujii Algorithm [4] using natural inclusion extension, T1: Ichida Fujii Algorithm using Taylor inclusion function [10], T 1 + M : Ichida Fujii Algorithm using Taylor inclusion function [10] with elimination of components by monotony, . RS: Optimization algorithm using the best right simplex, . RS+M: Method RS with elimination of components by monotony.
Pb NE T1 T1+M RS1 RS2 AS RS1+M RS2+M AS+M N 2832 731 185 329 271 319 101 59 110 f5 T(s) 30,99 17,31 1,81 3,58 3,00 3,09 1,06 0,62 1,71 C 312 24 24 24 18 22 24 12 22 N 322 421 108 228 25 228 210 25 34 f5 T(s) 1,43 11,88 1,11 2,36 0,28 3,05 2,05 0,28 0,72 C N 8 8 3959 2 468 8 861 2 564 8 878 8 302 2 205 2 309 f6 T(s) 201,86 9,07 19,31 12,18 22,25 6,73 4,48 10,05 C 216 77 137 101 121 86 67 78 N 3058 6483 6370 10221 2093 1786 2211 f7 T(s) 71,30 351,11 311,61 601,42 47,50 34,16 48,01 C 22 2934 2934 3012 65 17 117

. RS1: Optimization algorithm using right simplexes and intersection of the polyhedral cone with the unbounded cylinder, . RS2: Optimization algorithm using right simplexes and gradient relaxation, . AS: Optimization algorithm using the research of an admissible simplex, . RS1+M: Method RS1 with elimination of components by monotony, . RS2+M: Method RS2 with elimination of components by monotony, . A S +M: Method AS with elimination of components by monotony. (18) is not taken into account, the di erent methods are compared uniquely by lower bounds. These examples are sucient t o see the improvement provided by the algorithms 1, and 2 compared t o the classical Branch and Bound algorithm. Moreover the problem of clusters [2] is strongly reduced. The three di erent methods exposed here seem t o have quite the same eciency with perhaps a thin advantage for the algorithm using gradient relaxation. L Admissible simplexes have always the same initial vertex s = ( xL 1 , , xn ) which is surely not the best strategy. Secondly t o test the eciency of our enclosure methods, one computes the bounds of the function f 6 with decreasing width X1 X2 X3 X4 X5 X6 = = = = = = [1,1][1,1][1,1][1,1] [0,1][0,1][0,1][1,1] [0,1][0,1][0,1][0,1] [0,0.5][0,1][0,1][1,1] [0,0.5][0,0.5][0,1][1,1] [0 . 5 , 1] [0 . 5 , 1] [0 . 5 , 1] [ 0 , 1]

I n this example 0 G ( i X j ) , i = 1 , 2 , 3 , 4 and j = 1 , 2 , 4 , i n the others cases 0 G 4 ( X3 ), 0 G 1 ( X5 ) and 0 G i ( X6 ) , i = 1 , 2 , 3.

Boxes X1 Methods [ 3 , 29] f(Xi) NE [17 ,33] T1 [55 ,55] RS1 [44 ,82] RS2 [45.3,77.4] AS [32.2,55.7]

X2

X3

X4

X5

X6 [4,19] [1.5,24.5] [1.75,22.3] [3.8,19.3] [4,19] [3.8,19.3]

[1 .6 ,20] [0 .1 ,20] [12 ,28] [ 7 , 26] [24 ,32] [13.5,27.5] [7.4,28.6] [ 2 , 23] [7.4,28.6] [ 2 , 23] [7.8,28.1] [2.2,22.8]

[1 .6 ,16] [1 .6 ,16] [11 ,24] [ 9 , 21] [21 ,27] [19.1,24.9] [7.2,24.8] [ 6 , 22] [7.2,24.8] [1 .6 ,19] [7.6,24.4] [ 6 , 22]

Methods used t o compute upper and lower bounds are: . . . . f ( Xi ): range of f over X i , i = 1 , , 6 , NE: natural inclusion extension function, T1: Taylor inclusion function, RS1: right simplex and intersection of the polyhedral cone with the unbounded cylinder, . RS2: right simplex and gradient relaxation, . AS: research of an admissible simplex. First, we can see that our enclosure methods are not automatically better than classical natural inclusion extension of the function. However, when the width of the boxes decreases, our algorithms are more ecient and i n some cases, they are very accurate and they give the true bounds (method RS2 over box X 6 ).

Conclusion

The methods developed i n this paper, give a lower bound and/or an upper bound over a box X for a di erentiable multivariate function, and therefore may be used as an evaluation function, as well as enclosure methods of a global optimum. For a function dened over a box X of IRn , one estimation ofa minimum, for example, requires point evaluations i n n + 1 vertices of the box X and an interval evaluation of the partial derivatives. Eciency of algorithms is obtained by suitable choices of these vertices. Numerical results are very satisfactory on unconstrained problems; the rst results on optimization problems with constraints are promising.

References
1. G. ALEFELD Bounding the Slope of Polynomial Operators and some Applications, Computing : 26 (227-237) (1980). K. DU, R. B. KEARFOTT The Cluster Problem i n Multivariate Global Optimization , Journal of Global Optimization : (27-32) (1996). E. HANSEN Global optimization Using Interval Analysis, MARCEL DEKKER, Inc. 270 Madison Avenue, New York, New 10016 (1992). K. ICHIDA, Y. FUJII An Interval Arithmetic Method for Global Optimization , Computing : 23 (85-97) (1979). P. HANSEN, B. JAUMARD, J. XIONG Decomposition and Interval Arithmetic Applied to Global Minimization of Polynomial and Rational Functions, Journal Global Optimization : 3 (421-437) (1992). R. B. KEARFOTT Rigorous Global Search: Continuous Problems, Kluwer Academic Publishers, Dordrecht, Boston, London (1996). -

2.

10

3.

York

4. 5.

of

6.

7. J.L. LAGOUANELLE, F. MESSINE Algorithme dencadrement de loptimu m global dune fonction di erentiable , Compte Rendu `a lAcademie des Sciences, Numerical Analysis, T . 326, S erie I, (1998). 8. R.H. MLADINEO An Algorithm for Finding the Global Maximum of a Multimodal, Multivariate Function, Mathematical Programming : 34 (188-200) - (1986). 9. R. E. MOORE Interval Analysis, Prentice Hall, Inc. Englewood Cli s, N.J. (1966). 10. H. RATSCHEK, J. ROKNE New Computer Methods for Global optimization , ELLIS HORWOOD LIMI TED Market Cross House, Cooper Street, Chichester, West Sussex, PO19 1EB, England - (1988). 11. V . VISWESWARAN, C.A. FLOUDAS Uncontrained and Constrained Global Optimization of Polynomial Function i n one Variable ,JournalofGlobalOptimization: 2 (73-99) - (1992).

A T X macro package with JUCS style This article was processed using the L E