US Army Materiel Command AMCP 706-197 — AMCP 706-197 Reliability Prediction Handbook — US Army Engineering Design Guide PDF

Downloaded from https://www.everyspec.com tAMCP 74*l9l Y >O, and 0(y)-. for y O. If all con- straints g,(x) > 0 in Eq. 12-23 are satisfied, the summation term contributes nothing and co 0' - tat minimizing * is equivalent to minimizing f. If FIUE1-3any If is less than zero, io(gl) t which is s G certainly not anywhere near the minimum of s(x); thus, the summation term " penalizes TAt ct any violation of the constraints. uAsr proe- in dure which minimizes will never select a point outside the constrafint set and will, in The zig-zag motion shon here is time-consuming fact, select that point of the co2straint set and can be avoided by u Ring Zoutendijk's minimi- that minimizes f ei).

zation a frocedure. Unfortunately, there are certain difficul- ties that must be overcome in order to use FIGURE 12-13. An inefficient SrchProced re4 this powerful technique. To illustrate them, consider the problem of minimizing may + X2 subject to the constraint xa s 3; 12-3.5.2 Rosen's Gradient Projection Method x th(X., b). We fknow in advance that the solution to thi onbours o sx = 3, X b = 0. For At each iteration of Zoutendijkes proce- tdur example, dure, an optimization problem must be solved ~~~~to find a direction in which to move. AI- a()=2 + b X-3 •though this direction is in some sense "best", (12-24) ctthe procedure can be time-consuming.

An al- 1-.ative is provided by Rosen's Tcadient pro- Contours of d in uhe feasibn region (to the jection method, where a usable, feasible direc- right of the line x. = 3) are circles with :enter tion is found without solving an optimization at thie on an th3 penalty term 0o(x. - 3) problem (Ref. 16). This direction, however, has no effect Just to the left of (x d - 3), may not be loally "best" in any sense. bdn asowe Rosen's method, probably most efficient move to the left from x, = 3, we immediately when all constraints are linear, uses the cross all the contours of constant value.

A Kuhn-Tucker conditions both to generate new gradient minimization procedure starting at directions and as a stop criterion s Xa would move to the boundary at x and could proceed no further. In fact, since the unconstrained iNimumTof FNT ffunction is discontinuous and has no deriva- 12-3.6 NALT) +NTO TEC(2 tive along x = 3, minimization is almost NIQUEShopeless. 12-3.6.1 General These difficulties may be relieved by de- fining other, less "isthe" penalty fun tios.

Since powerful methods are available for or example, the function 01(y), where unconstrained minimization, it would seem 01 (y) = 0 for y> 0 and 01(y) =ky2 for convenient to solve constrained problems y 0, is continuous and has continuous first using unconstrained methods. This is exactly derivatives for all values of y, (k > 0). If 0, is what a "penalty function" alows us to do. used, the penalty for constraint violations is Instead of dealing with the constraints no longer infinite and somne violations are pos- Sdirectly, penalty function techniques find the sible.

unconstrained minimum of the function Coi)sider applying tis new penalty f, nc- ў(x)i = fx+ [g() (1-3 tion W the previous problem by minimizing whee ў.1is the penalty function, yet to be O(X)= =.2 + X2 +€(, ) determined. For example, suppose that the(125 penalty function is 0 0 (y), where 0 0 (y) = 0 for (12-25) 12-17