.. :,' Afrika Matematika Series 2 Vol. 4 (1991) . -. . o\~' , { \ . \ , M.A. IBIEJUGBA, T.A. ADEWALE and O.M. DAMIGDOLA Abstract It was generally believed, that because of the extensive work involved in coding and evaluating Hessian matrices, the most reasonable approach to unconstrained optimization that involves twice differentiable factorable functions is one which completely avoids these calculations. In this paper, by using the wellknown quasi-Newton method and the ingenuity of computer programming technique wielded to carry out direct inversion of the associated non-singular Hessian matrices of this class of problems, we obtain some numerical results which compare favourably with known results. 1. Introduction Like other ~econd deri~ative methods, a quasi-Newton method can contrive a quadratic approximation of the objectiverfu~ction ~nd .(uFthermore, the method has the distinct advantage or'quadr~tic, q:myergep~ejn, the vicinity of the minimum. Many experts in nonlinear ~ptl~i~atiori 'belfe~~ that second derivative inethods are more reliable because they usually require significantly fewer number of iterations and function evaluations than such methods which use only function or gradient informati~n: Hence this' explains the extreme popularity pf NE(~t~~. and quasi-Newtoll. methods in the, fielp, of unconstr~in~d nonlinear progra.mnung. One thing that has impeded nu~erical prpgre~s in n?nlinear programming is lack of a convenient method for representing and coding nonlinear functions of several variables, its gradient vector and. Hessian at a point that would make computer solutions easily available. McCormick [1] developed a factora.ble progra.rnm!ng c<;>de that has been helpful in this regard. For our subsequent development, we recall those definitions of factorable functions as defined in Refs. 2 and 4 as follows. 19 ~,',' '," 20 M.A. IBIEJUGBA et al. Definition 1.1: A function f( x 1, X2, . . . ,xn) is a factorable function of several variables if it can be represented as the last in a finite sequence of functions Uj(x)} which are composed as follows: (i) for j = 1, 2, .. ., n, fj(x) is given by fi(x) = xi; (ii) for j> n, fj(x) is of the form Ji,(x) + ft(x), k, e < j, or of the form Ji,(x)'. ft(x), k, e < j or of the form T[h (x)] , k < j, where T (f) is a function variable. of a single A quadratic approximation of the objective function, f( i), can be made by neglecting third-andhigher-order terms in the Taylor series expansion represented as follows: f(~) = f(~) + V'T(~)(~ - ~) + 1/2~(~ - ~fV'2 f(~)(~ - ~) ,. ... (1.1) where V'2 f(~) is the Hessian matrix, H(~), that is, the square matrix of second partial derivatives of the objective function, f(f..), evaluated at ~ and defined by o2f ox2 1 V'2 f(k) = H() = (~ - !.k) by -6x OXIOXn I I o2f OXnOXl Replacing o2f ... ... where 6Xk - !.k.,-l - !.k> (1.2) o2f ox then Equation (1.1) yields: f(~+J) = f(~) + V'T f(f..k)6xk ' 1/2'(6xkfV'2 f(~)6xk (1.3) Let the transition at kth iterative step from a point Xk to the point !.k+ 1 be defined by: !.k+ 1 ~+1 = !.k + 6Xk = !.k + >"k'!)k-= !.k + >"kPk' where, !.Ie = ycct;)[ :,'0;:1 ;;'ic ~O ;1:.\;+ l' P..k -" a unit, vector ir: the direction of l:-,Xk. (1.4 ) (1.5) MINIMIZING FACTORABLE FUNCTIONS 1!.1c 21 = any vector in the direction of ~Xk, ~k are chosen scalars such that, L::lxlc = .\Ic~ = >'IcEJc (1.6) The minimum of I(;!) in the direction of L::lxk is obtained by differt:ntiating I(;!) with respect to each of the components of L::lx and equati.ng the resulting expressions to zero to obtain: L::lxk = - [V2 1(~)r1 V I(~k), (1.7) where [.V2 I(xlc)] -1 is the inverse of the Hessian matrix of the objective function I(~, with respect to ~ at~. Putting Equation (1.7) into Equation (1.4) the transition from ~ to ~k+l for Newton's method becomes: ~+1 = ~ - V I(~)[V2I(xlc) ,1 -1 (1.8) It should be noted that both the direction and steplength are specified. If f(~) is actually quadratic, only one step is required to reach the minimum of I(i). However, for a general nonlinear objective function the minimum of I(~) will not be reached in one step, so that Equation (1.8) is usually modified to conform to the following equation: 1~+1 = ~ - ~/(~), (1.9) of the steepest descent method by introducing the parameter ..\ for the steplength into Equation (1.8) to give: Ak [V2 1(~k)r1 LV(~)] ~+1 = ~ - II [V2/(!.k)] 1 [V The ratio: (1.10) f(~)] II Ak "[I ['\72 1(~k)]=1 [Vl(;k)] II is just some scalar quantity which can be denoted by ~A:, hence, Equation (1.10) ~+1 or = ~kbecomes: - ~k [V2 1(::) r1 V I(Xk), - -1 ~+1 = -~k - AkH (~k)V I(~k) The search direction!!. is now given by Ek = -H-1(Xk)'\7/(~) (1.Ua) (1.Ub) (1.12) 22 ~tA\ IBIEtiUdBAet 8.1. 2:'q .Qua~iLN:~:wtbir!,~~t'hjj~ ja~g~ri~hm.' ,~,,:. , ."" ",,,,,, , " ,,'l," ",' )";'.'H~F,'n I'; .,' ',',( '~", n ,,'" : ""c',>"'" ," . 'r, > "".e, A ,p~!Ilber ,Qf,stJ.gg~t~C?JFh~~Y~ beep:off~red by many autho~s '(~ee, e.g: [2 ~ 10]) for 'the implementation 'Of Equation (1.11). Such suggestion§ include (i) ways of choOsing the de$ired steplength; (ii) suitable choice.of search directions; (iii) computational ways of avoiding the cumbersome task of coding and inverting th~ 'HeSsian matrix whei-e th~f iilverseexist'sand' (iv)1exploitation-strategies of expressing the 6uter product of the Hessian matrix. Most of the methods proposed can be classified into the following two categories: (i) 'negative curvature direction methods and (ii) descent-direction methods. However, several methods now combine features from both types. It is generally believed [2] that even when the Hessian matrix of the objective function is not positive definite, a lme search can be conducted &long a direction of negative curvature to reduce the value 'of the objective function. The fonowing steps are recommended by Polak [6; pp. 71 - 75] for solving such problems as min{f(~I~ E rn.n}, for !(~ that is twice continuously differentiable. ' Step 0: Select,; ~ E rn.nj select f3 E (0.5,0,8) and set k = O. Step 1: Compute, V!(?:,k), the gradient at ~. Step 2: If ~f(~k) = O,stop; else, go to step 3. Step 3: Compute H(%k) = a2 f(~)/ ail. the Hessian matrix at ~k' Step 4: If H1(Xk) exists, compute E(~k) by solving the equation H(!.AJE(fk) = - V !(~k), and go to step 5; else set Set>. = 1. Step 5: Step 6: Step 7: step 6. Step 8: 3. Compute 6 = EC~k) = -:- V !(~k) and go to step 5. >. f(~ + >',\:E(fk» - 2(Vf(~'\:)'E(~))' If 6::; 0, set >',\: = >. and go to step 8; else set >',\: = >'13 and go to Set ~+l = ~ + >'kE(~'\:); set k = k + 1 and go to step 1. Modifications of Newton's method Fiacco and McCormick [11] were the first to introduce the idea of finding a direction of negative curvature. Gill and Murray [10] developed a descent-direction method which avoids the eigen-vector analysis by performing an LDLT factorization of the Hessian matrix of the objective function. Emami and McCormick [12] developed a negative-curvature direction method which made use of the outer product form of the Hessian of a factorable function. Sisser [2] developed a 23. MINIMIZING FACTORABLE FUNCTIONS descent-direction method which also exploits the dyadic nature of the Hessian. Sisser perturbed the Hessian matrices by a quantity p added to the elements in th~ main diagonal where p. is determined to be a dyad. The resulting matrix is positive definite. A search parameter .A is included, so that the recursive formula for Newton's method then becomes: ~+1 = ~ - .Ak [H(~k) + Pk~-I ~i:) (3.1) Equation (3.1) can be rewritten as given below: v2 f = H == D + pI, where D is the diagonal, (3.2) 2m p. = L ISilJ Qt, (3.3) i=l+ I where the ai are n x 1 vectors and Si are scalars. The resulting Hessian matrix is inverted by the use of the Sherman-Morrison formula in [15] and generalized by Woodbury [13]: (A + asaT) -I = A-I - A-Ia (8-1 + aT A-Ia) -I aT A-I (3.4) In this study, a computer programme is written, to test for positive definiteness and to perform the task of inverting the Hessian matrix at each stage of the iteration. The experimental results of the authors of this paper on some test functions are given in Tables 1 - 12. More specifically, Table 1 is the minimization of Rosenbrock's Banana-shaped valley function, Table 2 is the minimization of Sisser's function, Table 3 is the minimization of Ros~nbrock's Cliff function, Table 4 shows the minimization of a power function while Table 5 is a summary of the results of the authors of this paper 'compared with those of Sisser's, etc. 4. A coordinate transformation It is usually profitable to scale the nonlinear programming problem whose solution is desired by optimization technique as this often has significant influence on the performance of the algorithm of interest. A well-scaled problem is one in which the contours of the objective function are approximately hyperspherical or elongated parallel to most search directions [3, pp. 30 - 34, 73 - 96, 190 - 217]. A good scaling device ensures speedy convergence of optimization routines and can lead to accurate solutions. Several effective scaling methods exist (see, e.g. Ref. 2) and could be used to computational advantages. One of such methods that involves the transformation of the variables from their physical nature to variables having properties that are desired in terms of optimization is employed here and 24 M.A. IBIEJUGBA et al. the computational experience of the authors of thi$ paper is clearly tabulated. For examplefj, Table 6 is the minimization.of transformed Rosenbrock's Bananashaped valley function, Table 7 is the minimization of Powell's quartic function, Table 8 shows the minimi~tion of hyperbola-circle function, while Table 9 is transformed Sisser's function, etc. Let p( Xl, X2, . . . , xn) be a point in a certain coordinate system. Consider a different coordinate system which can be generated from the original system by a simple rotation. Let the coordinates of the point p with respect to the new coordinate system be (Xl, X2, ..., Xn). Let the angle between the Xi-axis and the x;-axis be denoted by (Xi, X;); i, j = 1, 2,...2, n. Let f.i; be the direction cosine of the Xi-axis relative to the x;-axis. That is, f.u = cos(Xl, Xl), 62 = cos(Xl, X2), ..., 6n = cos(Xl, Xn) (4.1) Therefore, n (4.2) Xi = Lf.i;X;, i=I,2,...,n ;=1 The desired inverse transformation can be represented as (see [14]; pp. 1 - 16). n (4.3) xi=LE;iX;, i=I,2,...,n ;=1 It is convenient to arrange the Ei; into a square array called a transformation matrix J, as follows: Eu J = I El f.nl where, E12 ... 6n 62 ... E2n En2 ... Enn ( 4.4) n L Ei;Ek; = bik = {o, i= k (4.5) J=l 1, I = k are the orthogonality conditions obeyed and where of course, bik is the Kronecker delta. In this study, the test functions are transformed by suitable changes of variables and it is c:liscovered that in most cases the tranformation matrices (the Hessian matrices) of the resulting objective functions are diagonal matrices with positive real eigenvalues that guarantee positive definiteness. This is because the 25 MINIMIZING FACTORABLE FUNCTIONS transformations remove the cross-product terms of the quadratic approximation and the objective functions are transformed to the "Canonical forms" :3;. A val ley lies in the direction of the eigenvector associated with a small eigenvalue of the Hessian matrix of the objective function. The direction, - H( x) -1 V f( x), evaluated at different points in the space of the new variable always points toward the minimum of the objective function under investigation. 5. Numerical problems The following test problems are used: Example 5.1: Rosenbrock's Banana-shaped valley function, n = 2 f(XI,X2) = (1 - x.)2 + l00(X2 - xi)2, Starting point: (-1.2,1). Example 5.2: Powell's quartic function, n = 4, f(Xl, X2, X3, x.) = (Xl + lOx2)2 + 5(X3 -- x.)2 +- (X2 - 2X3). Starting - IO(XI - x.)', point: (3, -1,0,1). Example 5.3: Powell's badly scaled function, n = 2, f(Xl, X2) = (10.XIX2 - 1)2 +- [exp( -X.) + exp( -X2) - 1.000f2, Starting point: (0, 1). Example 5.4: Wood's function, n = 4, f(Xl,:Z:2,X3,X4) = l00(x2 - xi)2 + (1 - x.)2 -r- 9O(x4 - x5)2 +1O[(x2 - 1)2 + (x. - 1)2] + 19.8(x2 - l)(x. - 1), Starting point: (-3, -1, -3, -1). Example 5.5: Brown's badly scaled function, n = 2, f(Xl,X2) = (Xl - 106)2 + (X2 - 2.10-6)2 + (XIX2 - 2)2, Starting point: (1, 1). Beale's function, n = 2, Example 5.6: f(XI,X2) = [1.5 - Xl(1- X2W + [2.25 - xl(1 - x~W -r- [2.625 - xl(1 - x~W I Starting point: (1, 1). '- 26 M.A. IBIEJUG£A et al. Example 5.7: - Resenbrock's Cliff function, n = 2, l(xl,x2) = «XI - 3)/100)2 - (Xl - %2) + exp[20(xl - X2)], Starting point: (0, -"1). Example 5.8: Gottfried function, n = 2, I(xl, X2) = (xl - O.l136(XI + 3X2)(1 - Xl)]2 + [X2 + 7.5(2xI - x2)(1 x2)12, Starting point: (0.5, 0.5). Example 5.9: Hyperbola-circle function, n = 2, I(xl, X2) = (XIX2 - 1)2 + (x~ + x~ - 4)2, Starting point: (0, 1). Example 5.10: Sisser's function, n = 2, l(xl,x2) = 3x1- 2x~x~ + 3x~, point: (1, 0.1). Power function, n = 3t Example 5.11: Starting /(X1,X2, xs) '" It. ;xl Starting r ' point: (1, 1, 1). Example 5.12: A quadratic case, n = 3, I(Xl,X2,X3) = 1 + (XI + X2 + X3) + (x~ + 2x~ -j- 4x; - 2XIX2), Starting point: (0, 0, 0). The following transformations are suggested to scale some of the problems: Example 5.1a: Transformed Rosenbrock's function, n = 2, F(ZI,Z2) = zi + l00z~, where =) - XI, Z2 = X2 - XI' Starting point: (2.2, -0.44). 2 ZI 27 MINIMIZING FACTORABLE FUNCTIONS Powell's quartic function, n = 4, Example 5.2a: F(Zl,Z2,Z3,Z4) = zi -+ 5z~ -+ z~ + lOz:, where Zl Starting = xl + 10X2, Z2 = X3 - X4, Z3 = (X2 - 2X3)2, Z4 = (Xl - X4)2, point: (-7, -1, -1,4). Example 5.3a: Transformed Powell's badly scaled function, n = 2, F(Zl,Z2) = zi - z~, where Zl = 104x1X2 - 1, Z2 = [exp( -xd + exp( -X2) - 1.0001J ' Starting point: (-1, e-1 - 1.00(1). Example 5.5a: Transformed Brown's badly scaled problem, n - 22 F( Z2 - Zl +) Z2 + Zl Z2 + 10 Z2 - 2 . 1 Zl , [ 6 = 2, 0 -6 2 Zl, ] where Zl Starting = xl - 106, z2 = x2 - 2. 10-6, (1 - point: 106, 1 - 2.10-6). Example 5.6a: Transformed Beale's function, n = 3, F(zl, Z2, Z3) = zi + z~ + z;, where Zl = 1.5 - x1(1 - X2); Z2 = 2.25 - X1(1- xn; Z3 = 2.625 - x1(1 - x~); point: (1.5, 2.25, 2.625). Transformed Rosenbrock's Cliff function, n = 2, Example 5.7 a: Starting F(Zl, Z2) = 1O-4zi - Z2 + exp(20Z2), where, Zl = Xl - 3, Z2 = Xl - %2, Starting point: (-3, 1). Example 5.8a: Transformed Gottfried function, n = 2, 28 M. A. IBIEJUGBA et al. F( Zl, Z2) is obtainable by the substitution Starting point: (2, 0.5) Zl = Xl + 3X2, Z2 = 2X1 - X2. Example 5.9a: Transformed hyperbola-circle function, n = 2, F(Zl,Z2) is obtainable by the substitution Zl = X1X2 + 2, Z2 = Xl + X2, Starting point: (2, 1). Example 5.10a: Transformed Sisser's function, n = 2, F(Zl,Z2) is obtainable by the substitution Zl = xi, Zz = x~, Starting point: (1, 0.01). Example 5.11a: Transformed Power's function, n = 2, Consider the transformation Zi = x;, Starting point: (1, 1, 1). 6. Discussion of computatio~~esults {< The computer programme used for-the untransformed problems is slightly modified to suit the transformed version and run on model 128k Amstrad personal computer in single precision. The study has demonstrated that: (i) an algorithm cannot be expected to solve all problems but the robustness or reliability of an algorithm in obtaining an optimal solution for a wide range of problems can be tested experimentally by applying it to many test problems and from there a general comment can be drawn; (ii) a simple change in the starting point or initial step can have an impact on the search trajectory in minimizing an objective function [3]; (iii) when possible, a transformation of coordinates can be effected, thus spherising the contours of the objective functions and thus pointing the search direction toward the minimum. The algorithm failed for test Problem (5.2) but the result of the transformation suggested demonstrates that something can, at times, be done to get out of circumstances impeding the success of an algorithm on a test problem or a class. of problems. The failure of the algorithm on test Problem (5.2) can be ascribed to the fact that it is a quartic function with a Hessian matrix of rank 2. The function cannot therefore be approximated by a quadratic in the vicinity of the minimum as quasi-Newton method is understood to contrive. When transformed, the Hessian is reduced to a diagonal matrix with positive eigenvalues. However, the algorithm went through 30 iterations to bring the function under minimization to an accuracy of order 10-8 when gT g is of order 10-7. The stopping criterion is Ilgll < 10-6. This is an improvement over Sisser's result [1] reported as 10-5 for the same stopping criterion when the termination criterion is changed, the algorithm went through 158 iterations to bring MINIMIZING FACTORABLE FUNCTIONS 29 the function value to its minimum. One expects that after transformation the Rosenbrock's banana-shaped valley function would attain its minimum in much fewer iterations than (5.2) but geometrically the function represents a deep parabolic valley thereby necessitating more frequent changes of search directions; (iv) an algorithm can be rendered impotent by the way a problem is posed. Problems (5.3) and (5.5) demonstrate this. A total failure is met with transformed version of Gottfried function. Three different starting values are considered for Problem (5.7). The third starting point requires only six iterations to take the function to the minimum and with a better accuracy than the first two. For Problem (5.8) the accuracy is improved from the" order of 10-7 to 10-17 in twelve iterations. An unusual thing happened during the minimization of the power function as a result of Syivesters theorem that is coded into the programme to test for positive definiteness. There is a sudden change in the column indicating "positive definite" or "not positive definite". At the 42nd iteration, there is an indication of "not positive definite" after the minimum has been attained to the order of 10-27 in 40 iterations when gT 9 is zero. . Since some of the test problems are nonlinear functions that can be transformed into quadratic or an approximation to quadratic by changes of variables, it is therefore evident that conjugate gradient method can be employed to such cases. This will be the subject of the next paper. 30 M.A. IBIEJUGBA et al. TABLE 1 MbdmbaUon of Rosenbrock's banana-shaped valley function (untransformed) Iterative Step Function Value 0 2 4 6 8 24.2 4.10704765 2.75739953 1. 74224204 1.01702471 5.4 X 10-1 2.2 x 10-1 6.60 X 10-2 1.0 x 10-2 2.46 x 10-<1 2.57 x 10-0 8.08 x 10-16 0.0 10 12 14 16 18 20 21 *22 ,T, Xl 54227.3601 -1.2 911.214472 -9.2 x 10-1 364.647807 -5.7 10-1 127.668726 -2.4 x 10-1 43.7392975 -4.78 x 10-2 18.4492845 2.9 x 10-1 13.6092977 5.5 x 10-1 6.8860487 7.6 x 10-1 1.38608476 9.05 x 10-1 3.84 X 10-2 9.85 X 10-1 4.66 X 10-7 9.9 X 10-1 1.02 X 10-12 9.99 X 10-1 0.0 1.0 * Convergence in 22 iterations. Iterative Step Function Value X2 1.0 7.9 X 10-1 2.6 X 10-1 1.3 X 10-2 -3.09 X 10-2 6.5 X 10-2 2.9 X 10-1 5.6 X 10-1 8.2 X 10-1 9.7 X 10-1 9.9 X 10-1 9.99 X 10-1 1.0 TABLE 2 Minbnization of Sisser's function gTg Xl X2 0 2.9803 143.192144 1 6.1 2 1.16 x 10-1 1.10363149 4.4 x 10-1 4.4 X 10-2 4 4.5 X 10-3 8.5 X 103 1.9 X 10-1 1.9 X 10-2 1.77 X 10-<1 6 6.6 X 105 8.77x 10-2 8.77 x 10-3 10 2.69 X 10-7 3.89 X 10-9 1.7 X 10-2 1.7 X 10-3 12 1.05 x 10-8 3.0 X 10-11 7.7 X 103 7.7 X 10-4 14 4.1 x 10-10 2.3 X 10-13 3.4 X 10-3 3.4 X 104 16 1.6 x 10-11 1.78 x 10-15 1.5 x 10-3 1.5 X 10-4 18 6.2 x 10-13 1.37 X 10-17 6.7 X 10-4 6.7 X 10-5 .1.06 X 10-10 3.0 X 104 3.0 X 10-5 2.4 X 10-1<1 20 22 9.5 x 10-16 8.2 X 1022 1.3 X 104 1.3 X 10-5 24 3.7x 10-17 6.3 x 10-24 5.9 X 10-5 5.9 X 10-6 26 1.4 x 10-18 4.85 X 10-26 2.6 X iO-5 2.6 X 106 28 5.6 x 10-20 3.7 X 10-28 1.17 x 10-5 2.6 x 10-6 30 2.2 X 10-21 2.88 X 10-30 5.2 X 10-6 5.2 X 10-7 32 8.6 x 10-23 2.2 X 10-32 2.3 X 10-6 2.3 X 10-7 34 3.3 x 10-2<1 1.7 x 10-34 1.0 x 10-6 1.0 X 10-7 . Convergence in 40 iterations. 36 1.3 x 10-25 g - gradient 1.3 X 10-36 4.6 X 10-7 4.6 X 10-8 gT - transpose of the gradient vector. 38 5.1 x 10-27 ( 1.0 X 10-38 2.0 X 10-7 2.0 X 10-8 . 40 1.99 X 10-28 0.0 9.0 X 10-8 9.0 x 10-0 42 7.78 x 10-30 0.0 4.0 X 10-8 4.0 x 10-0 44 3.0 x 10-31 0.0 1.8 X 10-8 1.8 x 10-0 46 1.18 x 1032 0.0 7.94 x 10-0 7.94 X 10-10 48 4.6 x 10-3<1 0.0 3.5 x 10-0 3.5 X 10-10 50 1.8 X 10-35. 0.0 1.6 X 10-9 1.6 X 10-10 31 MINIMIZING FACTORABLE FUNCTIONS Iterative Step 0 1 2 3 4 6 6 7 8 9 10 11 12 TABLE 3 Mlnlmis.don 01 Roaenbrock 'a CUB' Function Function Value 1.0009 4.2 X 10-1 2'.6 X 10-1 2.1 X 10-1 2.0 X 10-1 2.0 X 10-1 2.0 X 10-1 2.0 x 10':"1 2.0 X 10-1 2.0 X 10-1 2.0 X 10-1 2.0 X 10-1 2.0 x 10-1 gTg 1.9999984 1.98 x 10° 1.43 x 10° 3.1 x 10- 1 5.2 X 10-2 1.18 X 10-'1. 2.8 X 10-3 6.9 x 10-. 1.7 x 10-. 4.3 x 1O1.1 x 1O2.6 X 10-6 9.0 X 10-13 XI 0 0 0 0 0 0 0 0 0 0 0 0 0 X2 1 4.2 X 10-1 2.4 X 10-1 1.75 x 10-1 1.6 X 10-1 1.5 x 10-1 1.5 X 10-1 1.5 X 10-1 1.6 X 10-1 1.5 x 10-1 1.49 X 10-1 1.49 X 10-1 1.49 X 10-1 TABLE 4 .._- -. - . --. ---..-- .. Iterative Step 0 KEYS: Function Value 7.11111111 gT, 707 .960618 XI 1 X2 1 3 2.8 x 10-1 6.46642091 2.9 x 10-1 2.9 X 10-1 6 2.14 X 10-3 3.69 X 10-3 8.77 X 10-'1. . 9 1.66 X 10-6 2.49 X 10-6 2.6 X 10-'1. 2.6 X 10-'1. 12 1.27 X 10-'7 1.69 X 10-11 7.7 x 10-3 7.7' x 103 16 9.79 x 10-10 1.14 X 10-1'1. 2.28 X 10-3 2.28 X 10-3 +ve 7.6 x P08itive 10-1'1. Definite 7.7 X 10-16 18 DEF 6.76 x 10. NOT +ve DEF Heaaian Not Positive Definite. 6.76 x 10-. 21 5.8 x 10-1. 6.2 X 10-111 2.0 x 10-. 6..76 x 10-. 24 4.48 x 10-16 3.54 x 10-'1.'1. 5.9 X 10-6 5.9 X 10-6 . 27 3.46 x 10-18 2.39 x 10-'1.6 1.76 X l(J6 1.76 X 10-5 . 30 2.66 X 10-'1.0 1.6 x 10-'1.8 5. X 10-6 5.2 X 10-6 S3 2.05 X 10-'1.2 1.09 x 10-31 1.5 X 10-6 1.6 X 10-6 S6 1.58 X 10-2. 7.4 X 10-36 4.6 X 10-'7 4.6 X 10-'7 39 1.2 x 10-26 5.0 X 10-38 1.3 X 10-'7 1.3 X 10-'7 42 1.27 x 10-28 0.0 4.3 x 10-8 4.3 x 10-8 - - Remarks +ve DEF . . . . . . H . . . NOT +ve DEF. M. A. IBIEJUGBA et al. 32 TABLE 5 c .'h 5i om parlSon 01 our resuns wu_- on-g'1"g Test Problem Author Number of Iterations 6.1 S 6'x.10 .-,;, 32 I 0.0 22 S 5.2 x 10 .", 31 I 4.0 x 10 ." 15 S 2 x 10 ., I 1.7 x 10 .,,, S 1.0 x 10 ." I 1.3 (F) 6.5 S 2 x 10 '4'> I 1.0 (F) 6.6 S 2 x 10 .-,;, I 0.0 1 6.7 S F F I 9 x 10 6.8 S 4 x 10 ." (F) 6.2 6.3 6.4 141 3 32 F 12 F 13 12 13 F I ." S - Sisser; I - Ibiejugba et1.5 al.;xF10-Algorithm failed 6.9 S 9 x 10 .;>u 6.10 12 9 I 5.5 x 10 " 12 S -TABLE 6 22 I 0.0 valley function (transformed) Minimisation of Rosenbrock's banana-shaped 39 Iterative! 0 1 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 51 ( Step I Function Value 24.2 6.05 3.78 x 10-1 5.90 x 10-3 9.2 X 10-5 1.44 x 10-6 2.25 x 10-8 3.5 x 10-10 5.5 x 10-12 8.6 x 10-14 1.3 x 10-15 2.09 X 10-17 3.3xlO-19 5.1 x 10-21 8.0 x 10-23 1.25 x 10-24 1.9 x 10-26 3.0 x 10-28 4.7 x 10-30 gTg 7763.36 1940.84 121-.3025 1.89535156 2.96 X 10-2 4.60 X 10-4 7.2 X 10-6 1.13 X 10-7 1.76 X 10-9 2.76 X 10-11 4.3 X 10-13 6.7 x 10-15 1.05 x 10-16 1.6 X 10-18 2.57 X 10-20 4.0 X 10-22 6.27 X 10-24 9.79 X 10-26 1.53 X 10-27 Xl 2.2 1.1 2.75 x 10-1 3.4 x 10-2 4.29 X 10-3 5.37 x.1O-4 6.7 X 10-5 8.39 X 10-6 1.05 X 10-6 1.3 X 1C,-7 1.64 X 10-8 2.0 X 10-9 2.6 X 10-10 3.2 X 10-11 4.0 X 10-12 5.0xlO-13 6.25 X 10-14 7.8 X 10-15 9.7xlO-16 X2 -0.44 -0.22 -0.055 -6.9 X 10-3 -8.59 X 10-4 -1.07 x 10-4 -1.34 X 10-5 -1.67 X 10-6 -2.09 X 10-7 .-2.6 X 10-8 -3.28 x 10-0 -4.09 X 10-10 I -5.1 X 10-11 -6.4 X 10-12 -8.0 X 10-13 -1.0 X 10-13 -1.25 X 10-14 -1.56 X 10-15 -1.95 X 10-16 MINIMIZING FACTORABLE FUNCTIONS Iterative Step 0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96 102 108 114 120 126 132 138 144 150 156 162 Function Value 33 TABLE 7 MiniInization of Powell's quartic function (transformed) g7.'g Xl X2 X3 X4 215 5.026035 6700 21.4811359 -7 -2.2 x 10-0 -1 6.8 X 10-2 1 3.15 X 10-2 4 -5.3 X 10-2 9.89 X 10-2 7.13 X 10-1 -2.97 X 10-1 -3.7 X 10-3 4.2 X 10-2 2.7 x 10-3 6.4 x 10-5 1.6 X 10-6 4.2 x 10-8 1.8 x 10-0 5.08 x 10-11 9.9 x 10-13 2.7 x 10-14 6.15 x 10-:-16 1.56 x 10-17 6.47 x 10-19 1.8 x 10-20 3.6 X 10-22 9.9 x 10-24 2.3 x 10-25 5.9 x 10-27 2.3 x 10-28 6.6 x 10-30 1.3 x 10-31 3.6 x 10-33 8.5 x 10-35 2.2 x 10-36 9.96 x 10-38 9.3 x 10-30 0.0 1.7 X 10-2 3.1 x 10-4 2.4 x 10-5 5.1 X 10-7 7.6 x 10-0 2.1 X 1010 6.6 X 10-12 1.6 X 10-13 9.87 X 10-15 2.09 X 10-16 2.77 X 10-18 7.7 X 10-20 2.6 X 10-21 6.2 X 10-23 4.1 X 10-24 8.6 X 10-26 1.0 X 10-27 2.78 X 10-20 1.0 X 10-30 2.39 X 10-32 4.2 X 10-34 3.5 X 10-35 6.8 X 10-37 1.55 X 10-37 8.9 x 10-38 -4.9 X 10-2 -7.8 X 10-3 -1.06 X 10-3 -1.77 x 10-4 -4.2 x 10-5 -7.0 X 10-6 -9.5 X 10-7 -1.6 X 10-7 -2.0 X 10-8 -3.4 x 10-0 -7.9 X 10-10 -1.3 X 10-10 -1.8 X 10-11 -3.0 X 10-12 -3.8 X 10-13 -6.4 x 10- I'll -1.5 x 10-14 -2.5 X 10-15 -3.4 X 10-16 -5.7 X 10-17 -8.9 X 1Qt8 -1.2 X 10-18 -3.16 X 10-19 -9.6 X 10-20 -2.3 X 10-20 -5.57 x 10-5 -4.5 x 10-5 2.47 X 10-6 3.69 X 10-7 5.0 X 10-9 7.5 X 10-10 -4.1 X 10-11 -6.1 X 10-12 1.8 X 10-13 2.7 X 10-14 3.7 X 10-16 5.5 X 10-17 -2.9 X 10-18 -4.5 X 10-19 1.3 X 10-20 1.9 X 10-21 2.7 X 10-23 4.0 X 10-24 -2.2 X 10-25 -3.3 X 10-26 -2.7 X 10-27 1.45 xlO-28 4.46 X 10-30 4.3 X 10-31 -1.6 X 10-32 2. 2.9 x.lO4.2 X 10-3 -3.9 x 10-4 2.2 x 10-4 3.1 x 10-5 -1.1 X 10-6 -1.5 X 10-6 8.6 X 10-8 1.2 X 10-8 4.5 x 10-0 6.4 X 10-10 -2.2 X 10-11 -3.1 X 10-12 1.7 X 10-12 2.5 X 10-13 9.3 X 10-14 1.3 X 10-14 -4.6 X 10-16 -6.5 X 10-17 3.6 X 10-17 5.14 X 10-18 -4.8 X 10-19 2.7 X 10-19 2.6 X 1O-2C1 -1.7 X 10-20 1.49 X 10-20 . 7.1 X 10-3 1.12 X 10-3 1.5 x 10-4 2.5 x 10-5 5.97 X 10-6 1.0 X 10-6 1.35 X 10-7 2.28 X 10-8 2.86 X 10-9 4.8 X 10-10 1.1 x 1O-1C1 1.9 X 10-11 2.6xlO-12 4.3 X 10-13 5.4 X 10-14 9.1 X 10-15 2.15 X 10-15 3.6 X 10-16 4.86 X 10-17 8.18 X 10-18 1.28 X 10-18 1.7 x 10-19 4.5 X 10-20 1.4 X 1O-2C1 3.3 X 1021 TABLE 8 Minimization of Hyperboia-circle function (transformed) Function Value g'g Xl X2 0 10 244 2 1 3 2.1 X 10-1 3.15 x 100 2.5 2.3 6 3.06 x 10-3 7.3 X 10-2 2.9 2.4 Iterative Step 9 4.7 x 10-5 1.19 X 10-3 2.99 2.447 12 7.38 x 10-7 1.87 x 10-5 2.999 2.449 15 1.15 x 10-8 2.9 X 10-7 2.9999 2.4494 18 1.18 x 10-10 4.6 X 10-9 2.99998 2.44948 21 2.8 x 10-12 7.15 X lOll 2.999984 2.449489 24 4.4 x 10-14 1.12 X 10-12 2.9999998 2.4494896 27 7.11 x 10-16 2.06 X 10-14 2.99999998 2.4494897 30 7.8 x 10-18 3.1 X 10-17 3 I 31 8.67 x 10-19 3.46 X 10-18 3 2.44948974 . Convergence in 31 iterations. 2.44948974 . i - ~--- 34 M.A. IBIEJUGBA et al. TABLE 9 Minimisation of transformed Slaser's funed Iterative Step 0 * 1 2 Iterative Step 0 1 2 3 4 5 6 7 8 9 10" 11 12 .. * Starting Xl u.. X2 2.98 40.004 1 0.01 3.57 X 1O-2 4.76 X 10-21 0 -1.09 X 10-11 0 0 0 0 * There is convergence in one iteration. Function Value 5.78992996 2.06310453 1.23835341 1.18578523 1.17372824 1.16980327 1.1698031 1.16980306 1.1698036 1.1698036 1.16980306 1.16980306 1.16980306 TABLE 10 Minimisation of Gottfried function g'g 2295.81781 258.38057 25.45483 7.42725042 1.94669061 2.84660 x lO-. 7.25414 x 10-f> 1.85 x 10-5 0.603005063 4.71998 x lO-6 1.20479 x 10-6 9.25225 x 10-7 9.25225 x 10-7 9.25225 x 10-7 Xl X2 0.5* 0.516210901 0.544568287 0.567457805 0.584203534 0.602809815 0.602938802 1.04023337 0.603039116 0.603056619 0.603058869 0.603058869 0.603058869 0.5* 0.71657717 0.895810376 0.958387374 0.997670211 1.03983364 1.04009894 1.04030151 1.04033607 LO034045 1.04034045 1.04034045 point given. Convergence in 10 iterations. Iterative Step 0 1 2 3 4 5 6 7 8 9 .. g"-g Function Value TABLE 11 Minimisation of Gottfried function Function Value 1.09072144 1.16985437 1.1698154 1.16980601 1.16980376 1.16980322 1.1698031 1.16980306 1.16980306 1.16980306 10 1.16980306 11 1.16980306 * 12 1.16980306 9:9 Xl X2 4.2523169 5.94087 x 10-3 1.46699 x 10-3 3.62374 x 10-4 8.95341 x 10-5 2.21258 x 10-"1> 5.46878 x 10-6 1.35197 x 10-6 3.34306 x 10-7 1.20414 x 10-10 9.33451 x 10-14 2.01716 x 10-16 1.48205 x 10-17 0.5** 0.604701673 6.0 X lO-1 6.0 X 10-1 6.0 X lO-1 6.0 X lO-1 6.0 X lO-1 6.0 X 10-1 6.0 X 10-1 6.03075573 6.0307515 6.03075138 6.03075138 1** 1.03961707 1.0 1.0 1.0 1.0 1.0 1.04036848 1.040368 1.04037207 1.04037165 1.04037 1.04037164 * Convergence in l~ iterations. Starting point changed: improved accuracy. 35 MINlw.:3ING FACTORABLE FUNCTIONS TABLE 12 .Jmparison ~f results with transformed problems versus original problems Test problem 6.1a 6.2a 6-3a 6.5a 6.6a T 0 T 0 T 0 T 0 T Min. Fund. Value 0.0 0.0 0.0 F 1.2 x 10 -.. 2.0 0.0 0.0 0.0 g.1"g 0.0 0.0 2.7 x 10 ..>1 F 4.9 x 10. 1.7 x 10 -,,, 0.0 0.0 0.0 No. of Iterations 52 22 158 F 4(F) .2 1 2 1 0 1.4 x 10' 0.0 1 T 1.9 x 10 -, 2 x 10 ., 2.3 x 10.w 2.7 x 10 .,,, 26 24 6.7a _. 2 x 10-1 9 X 10-13 2xl0-10 1 X 10-10 6 T F 1.17 F 9.2 x 10 F 10 0 1.17 1.5 x 10-17 12 0 . 6.8a T - Tn..::::>rmed Problem; .. 0- 12 Orical Problem; T 8.7 x 10 .1" 3.5 x 10 .1" 31 5.5 x 10 .'1 12 F - Th, ."orithm failed. 6.9a 0 1.3 x 10'0 T 0.0 0.0 . Chao..:>f starting point improves both minimum function value and gT2g. 6.lOa 0gT g.0.0 Ch"-=e of starting point improves T 1.26 x 10 -.... Refe:-'!1ces 6.l1a 0 2.4 x 10-"0 0.0 0.0 0.0 .. 40 17 40 1. _? McCormick, (1974): A minimanual for the lJ6e of the SUMT Computer Program :;d the Factorable Programming Language, Standford University, Systems Optimization _:Joratory, Technical Report No. SOL - 74 - 15, 1974. 2. - .. Sisser, (1982): A modified Newton'lS method for minimizing factorable functionlS, Jour :.. of Optimization Theory and Applications, Vol. 38, No.4, Dec. 1982, pp. 461 .~ 3. - .-'1. David, (1972) Applied nonlinear programming, McGrawhill Book Company, New -. ~k, 1972. 4. - 3. Gerald, (1980): Applied numerical analYlSilS, Addison-Wesley Publishing Company, ::,ading, Massachusetts, 1980. 5. . -::;. Singh and A. Titli, (1978) SYlSterru optimization and decompolSition, Pergamon Press, - 78. 6. : Polak, (1971): Computational methods in optimization, Academic Press, New York, -71. 36 M.A. IBIEJUGBA et al. 7. D. RU88ell, (1970): Optimization theorl/, W.A. Benjamin, Inc., New York, 1970. 8. P.E. Gill and W. Murray, (1978): Algorithml lor the .olution of lea.t .quaru probleml, SIAM Journal of Numerical Analysis, Vol. 15, No.5, Oct. 1978, pp. 977 - 991. 9. J.C. Nash, (1979): Compact numerical metholblor compute", linear algebra and function minimization, Adam Hilger Limited, Bristol, 1979. 10. P.E. Gill and W. Murray, (1974): Newton tl/pe method. lor unco"'trained linearly con .tructed optimization, Mathematical Programming, Vol. 7, 1974, pp. 311 - 350. 11. A.Vi. Fiacco and G.P. McCormick, (1968): Nonlinear programming: sequential uncon .trained minimization technique., John Wiley and Sons, New York, 1968. 12. G. Emami and G.P. McCormick, (1978): U.e 01 a .table generalized inver3e algorithm to evaluate Newton method "rategie., The George Washington University, Institute for Management Science and Engineering, Program in Logistics, Technical Report Serial No. T-384, 1978. 13. M. Woodbury,(1950): Inverting modified matrice., Princeton University, Princeton, New Jersey, Statistical Research Group, Memorandum No. 42, 1950. 14. J.B. Marion, (1970): Cl~.ical dl/namic. of particle. and .y&tem8, Academic Press, New York, 1970. 15. J. Sherman and W.J. Morrison, (1949): Adjwtment of an inverse matrix corresponding to change. in the element, of a given column or a given row of the original matrix, Annals of Mathematical Statistics, Vol. 20, pp. 621, 1949. M.A. IBIEJUGBA, T.A. ADEWALE Department of Mathematics, University of Ilorin, Borin, Nigeria and O.M. BAMIGBOLA,