Downloaded from https://www.everyspec.com AMCP 706-197 The contours of this function to the right of ing point depends, of course, on the choice of x. = 3 are circular but to the left they are r,, i nd is denoted by x(r). By this reasoning, elongated ellipses, showing the same bunching x(r 1 ) will always be inside the constraint set. effect as before. This effect gets worse as k If this minimization process is repeated increases. for a sequence of values r 1 > r 2 > ... r k > 0, A gradual approach is more practical. each minimizing point x(r , ) also will be strict- Rather than solve only one unconstrained ly inside the constraint set.
Furthermore, as problem, we solve a sequencD of such prob- the value of r is reduced, the influence of the lems, each one bringing us closer to the final term which "penalizes" closeness to the con- solution. For example, we can solve the prob- straint boundaries (the last term in Eq. 12-26) lem with a small value of k. Then, using that also is reduced and, in minimizing* (xr), solution as a starting point, choose a larger more effort is concentrated on reducing the value of k and re-solve the problem. Repeat f(x) term. Thus, the sequence of points the procedure several times.
In general, the x(r, ),x(r 2 ),- -- can come as close as necessary sequence of unconstrained minima ap- to the boundary of the constraint set. We proaches the solution of the original con- would expect that as r approaches zero, the strained problem. minimizing point x(r) approaches the solution When the penalty function 01 is used, of the original problem of minimizing fx) intermediate solutions usually violate the con- bubject to the constraintsg, > 0. straints. Thus, the method approaches the This method is particularly attractive in constrained minimum from outside the con- dealing with problems that have markedly straint set.
In many -cases, this may be unsatis- nonlinear constraints, since it approaches the factory. If small violations of the constraints solution value from inside the constraint set. are not permitted, intermediate solutions Motion along the boundaries of this set, - often cannot be used. The method is ineffi- which can be very cumbersome when the cient if the objective or constraint functions boundaries have large curvature, is completely are ill-behaved exterior to the constraint set. avoided.
Moreover, the approach cannot be used at all Fiacco and McCormick have shown that when any of these functions is not defined all the previous conjectures are true under cer- outside of the constraint set. tain conditions (see Table 12-2). Condition 7 is not implied by conditions 5 and 6, but only 12-3.6.2 The Fiacco-McCormick Method small additional requirements on f and g, are needed for it to hold (Ref. 16). The Fiacco-McCormick method avoids the The Fiacco-McCormick procedure is given difficulties we just described by approaching in Table 12-2. Step 2 may be accomplished by the optimum from inside the constraint set any of the unconstrained minimization proce- (Refs. 17 and 18).
To use this method, we dures in this paragraph. In Step 3, r ought to first define the function be reduced by dividing each time by the same in factor. YT'(x, r) =f(x)+ r 1 Z g,( W (12-26) 124 DYNAMIC PROGRAMMING where r > 0. Let r 1 -* 0 and choose x 0 inside Dynamic programming is a general ap- the constraint set. In the problem of minimiz- proach for solving a sequential decision proc- ing I(x,r 1 ) starting from x 0 and subject to no ess. Optimization is merely one kind of se- constraints, a minimum must exist inside the quential decision process. This topic is not constraint set, since *I,(,r 1 ) .
o on the grasped easily from a short exposition, nor is boundary of this sell (because some g(x) = 0). it often practical for reliability problems, Thus, the path of steepest descent leading except when the problems can as easily be from the point x 0 (a path on which I(X,r 1 ) is solved another way. Therefore, several refer- strictly decreasing) cannot penetrate the ences (R9fs. 19-22, 34) are gi' en for further boundary of the constraint set, The m.nimiz- study, should the need arise. 12-1a