You are on page 1of 2

94

applicable for large values of n is K n+1 Fn 5 (4.24)

(e.g., if n = 11, Fn = 1.44 and K n+1 / 5 144.001). Thus Eqs. (4.5) and (4.24) give the region of uncertainty for the Fibonacci search as 5 I1 n+1 I1 F = In = Fn K Similarly, for the golden-section search GS = In = and hence GS K2 = 1.17 F 5 Therefore, if the number of iterations is the same in the two methods, the region of uncertainty in the golden-section search is larger by about 17% relative to that in the Fibonacci search. Alternatively, the golden-section search will require more iterations to achieve the same precision as the Fibonacci search. However, this disadvantage is offset by the fact that the total number of iterations need not be supplied at the start of the optimization. An implementation of the golden-section search is as follows: Algorithm 4.2 Golden-section search Step 1 Input xL,1 , xU,1 , and . Step 2 Assign I1 = xU,1 xL,1 , K = 1.618034 and compute I2 = I1 /K xa,1 = xU,1 I2 , xb,1 = xL,1 + I2 fa,1 = f (xa,1 ), fb,1 = f (xb,1 ) Set k = 1. Step 3 Compute Ik+2 = Ik+1 /K I1 K n1

One-Dimensional Optimization

95

If fa,k fb,k , then update xL,k+1 , xU,k+1 , xa,k+1 , xb,k+1 , fa,k+1 , and fb,k+1 using Eqs. (4.7) to (4.12). Otherwise, if fa,k < fb,k , update information using Eqs. (4.13) to (4.18). Step 4 If Ik < or xa,k+1 > xb,k+1 , then do: If fa,k+1 > fb,k+1 , compute x = 1 (xb,k+1 + xU,k+1 ) 2 If fa,k+1 = fb,k+1 , compute
1 x = 2 (xa,k+1 + xb,k+1 )

If fa,k+1 < fb,k+1 , compute


1 x = 2 (xL,k+1 + xa,k+1 )

Compute f = f (x ). Output x and f , and stop. Step 5 Set k = k + 1 and repeat from Step 3.

4.5

Quadratic Interpolation Method

In the approximation approach to one-dimensional optimization, an approximate expression for the objective function is assumed, usually in the form of a low-order polynomial. If a second-order polynomial of the form p(x) = a0 + a1 x + a2 x2 (4.25) is assumed, where a0 , a1 , and a2 are constants, a quadratic interpolation method is obtained. Let (4.26) p(xi ) = a0 + a1 xi + a2 x2 = f (xi ) = fi i for i = 1, 2, and 3 where [x1 , x3 ] is a bracket on the minimizer x of f (x). Assuming that the values fi are known, the three constants a0 , a1 , and a2 can be deduced by solving the three simultaneous equations in Eq. (4.26). Thus a polynomial p(x) can be deduced which is an approximation for f (x). Under these circumstances, the plots of p(x) and f (x) will assume the form depicted in Fig. 4.9. As can be seen, the minimizer x of p(x) is close to x , and if f (x) can be accurately represented by a second-order polynomial, then x x . If f (x) is a quadratic function, then p(x) becomes an exact representation of f (x) and x = x . The rst derivative of p(x) with respect to x is obtained from Eq. (4.25) as p (x) = a1 + 2a2 x

You might also like