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 AMP706-157 TABLE 12-1 (cond) OPT IZING UNCONSTRAINED PROBLEM .Comput anew vector o hoWs IAod 4. If f 3 > f, or if (f - 2 t2 + 1. Forr-l, 2,..., n, calculate f)(f,-f2-A)Z >%A. (r,- 4. Compute V f(i+ 1 ). mum (se cubic or quadratic inter- d -, 1. for the next iteration and usne i, as the next 3o. 5. Compute poisdon pcedure) 5. If neither definteon in step _ so that f(,,fs), or both, use the od dirctions B-Vfi,+ 1 ) v61+ 1 ) XT" Xrl +0r 4 had,, defne - (. -io), ad B V r alf'( ) V f j) 2. Find the. integer m, 14 m c alcul a so tht fx- + ais) is a 6.

Compute a succesor direc- n, so that [f (,.,) -f(i,)] ja inimam (we cubic or quadratic tion, maximum, and define interpolation procedure). 6.usiii 5+1 =-V f(X,+a 1 . A f. 1 )-f(7) . as the directions for the next 7. Check a stop criteri. If -1 iteraioand i. = for the next s ztop 3. Calculate f - f(. - o) - is satisfied, stop. Otherwise, r t.-ddfn f (o adf O tosdep2andreplaeeib-i+l, t~ d defnef ( 0 ) and (2 " 6 " . frx.). 6 fit a cubic polynomial to the computed contours are circles (or, in the n-dimensional values g(.), g.), g(,. and 9;b). This poIyno- case, hyperspheres), the method finds the mial has a unique minimum located at a. in minimum in one step.

However, for other the interval between a and b. In Step 7, ct, = contours, the gradient direction is generally taken as the desired value of a* if ar is a quite different from the direction to the mini- better choice than either a or b. If not, the mum, and the method produces the ineffi- interpolation is repeated over a smaller inter- cient zig-zag behavior shown in Fig. 12-1. val in Steps 8 and 9.

Since many, if not most, of the functions If derivatives are not available or are diffi- occurring in practical applications have eccen- cult to compute, the quadratic interpolation tric or nonspherical contours, we often must procedure can be used for 1-dimensional mini- turn to more efficient methods than steepest mization. Step 5 of this procedure fits a quad- descent. ratic polynomial to the three values g(.), gI (b , and g,. The minimum of this polynomial is located at a.

12-2.2 SECOND-ORDER GRADIENT The most that can be guaranteed by the METHODS steepest descent method, or any other itera- tive minimization technique, is that it will A number of minimization techniques find a local minimum, usually the one "near- have been developed to overcome the difficul- est" to the starting point x 0 . To attempt to ties of the method of steepest descent. The find all local minima (and thus the global min- general notion behind these techniques is that imum), the usual approach is to repeat the methods which quickly and efficiently mini- minimization from many different initial mize a general function must fulfill two crite- points. ria.

They must work well on a quadratic func- tion, and they must be guaranteed to con- 12-2.1.3 Numerical Difficulties verge (eventually) for any general function. These criteria are based on the observation Since successive steps of the method of that, since the first partial derivatives of a steepest descent are orthogonal, some func- function vanish at the minimum, a Taylor tions converge very slowly. If the function series expansion about the minimum x* yields 12-4