14. Nonlinear least-squares

Transcription

1 14 Nonlinear least-squares EE103 (Fall ) definition Newton s method Gauss-Newton method 14-1

2 Nonlinear least-squares minimize r i (x) 2 = r(x) 2 r i is a nonlinear function of the n-vector of variables x r(x) = (r 1 (x),r 2 (x),,r m (x)) reduces to linear least-squares if r(x) = Ax b g(x) = r(x) 2 may have multiple local minima; finding the global minimum is usually very hard a nonlinear minimization problem; can apply Newton s method Nonlinear least-squares 14-2

3 Interpretation as overdetermined nonlinear equations m nonlinear equations in n variables r 1 (x 1,x 2,,x n ) = 0 r 2 (x 1,x 2,,x n ) = 0 r m (x 1,x 2,,x n ) = 0 usually there is no x that satisfies r(x) = 0 instead we can calculate the vector x that minimizes r(x) 2 = r i (x) 2 Nonlinear least-squares 14-3

4 Inductor modeling example 50 nonlinear equations in 5 variables x 1,, x 5 exp(x 1 )n x 2 i wx 3 i d x 4 i Dx 5 i L i, i = 1,,50 method 1 (exercise 84): suppose we are free to choose the error criterion if we choose to minimize the sum of squared errors on a logarithmic scale, minimize 50 (log(exp(x 1 )n x 2 i wx 3 i d x 4 i Dx 5 i ) logl i ) 2, we obtain a linear least-squares problem: minimize Ax b 2 where A = 1 logn 1 logw 1 logd 1 logd 1 1 logn 2 logw 2 logd 2 logd 2 1 logn 50 logw 50 logd 50 logd 50, b = logl 1 logl 2 logl 50 Nonlinear least-squares 14-4

5 method 2: minimize the sum of squared errors on a linear scale minimize 50 (exp(x 1 )n x 2 i wx 3 i d x 4 i Dx 5 i L i ) 2 this is a nonlinear least-squares problem: minimize 50 r i (x) = exp(x 1 )n x 2 i wx 3 i d x 4 i Dx 5 i L i r i (x) 2 where = exp(x 1 +x 2 logn i +x 3 logw i +x 4 logd i +x 5 logd i ) L i much harder than linear least-squares (may have multiple local minima) requires an iterative method can use method 1 to find a starting point Nonlinear least-squares 14-5

6 Navigation from range measurements estimate position (u,v) from distances to m beacons at locations (p k,q k ) measured distances: ρ i = (u p i ) 2 +(v q i ) 2 +w i, i = 1,,m where w i is range error, unknown but small nonlinear least-squares estimate: choose estimates (u, v) by minimizing g(u,v) = = r i (u,v) 2 ( ) 2 (u pi ) 2 +(v q i ) 2 ρ i Nonlinear least-squares 14-6

7 example correct position is (1, 1) ( ) m = 5 beacons at positions range measurements accurate to ±02 graph of g(u,v) 4 contour lines of g(u,v) u v 4 v (global) minimum at (118, 082) local minimum at (299, 212), local maximum at (194, 187) u Nonlinear least-squares 14-7

8 Newton s method for nonlinear least-squares apply method of page to g(x) = r(x) 2 = m first and second derivatives of g: g(x) x k = 2 2 g(x) x j x k = 2 r i (x) 2 r i (x) r i(x) x k ( r i (x) 2 r i (x) + r ) i(x) r i (x) x j x k x j x k ie, the gradient and Hessian of g are g(x) = 2 2 g(x) = 2 r i (x) r i (x) ( ri (x) 2 r i (x)+ r i (x) r i (x) T) Nonlinear least-squares 14-8

9 example (inductor problem of page 14-4) method 1 (linear least-squares): x 1 = 725, x 2 = 138, x 3 = 048, x 4 = 028, x 5 = 121 method 2: apply Newton s method to g(x) = 50 (exp(x 1 +x 2 logn i +x 3 logw i +x 4 logd i +x 5 logd i ) L i ) 2 use as starting point the linear least-squares solution converges in four iterations to x 1 = 714, x 2 = 132, x 3 = 053, x 4 = 022, x 5 = 127 Nonlinear least-squares 14-9

10 example (navigation problem of page 14-6) v u from starting points (0,0), (0,4), (4,0), converges to 1 (minimum) from starting points (4, 4), (2, 2), converges to point 2 (local minimum) Nonlinear least-squares 14-10

11 convergence from starting point x (0) = (15,4) x (k) x v x (0) x (1) k u converges to minimum at (118, 082) 2nd and 3rd iteration use negative gradient direction; other iterations use Newton direction Nonlinear least-squares 14-11

12 Gauss-Newton method for nonlinear least-squares a simpler alternative to Newton s method for minimizing g(x) = r i (x) 2 start at some initial guess x (0) at iteration k, linearize r i (x) around current guess x (k) : r i (x) r i (x (k) )+ r i (x (k) ) T (x x (k) ) new guess x (k+1) is the minimizer of ( ) 2 r i (x (k) )+ r i (x (k) ) T (x x (k) ) ie, sum of the squares of the linearized residuals Nonlinear least-squares 14-12

13 to find x (k+1) from x (k), solve a linear least-squares problem minimize A (k) x b (k) 2 with A (k) = r 1 (x (k) ) T r 2 (x (k) ) T r m (x (k) ) T, b (k) = r 1 (x (k) ) T x (k) r 1 (x (k) ) r 2 (x (k) ) T x (k) r 2 (x (k) ) r m (x (k) ) T x (k) r m (x (k) ) advantage (over Newton s method): no second derivatives of r i needed disadvantage: convergence is slower Nonlinear least-squares 14-13

14 Summary: Gauss-Newton method given initial x, tolerance ǫ > 0 repeat 1 evaluate r i (x) and r i (x) for i = 1,,m, and calculate r := r 1 (x) r 2 (x) r m (x), A := r 1 (x) T r 2 (x) T r m (x) T 2 if 2A T r ǫ, return x 3 x := (A T A) 1 A T b until maximum number of iterations is exceeded, b := Ax r in step 2, note that 2A T r = g(x) Nonlinear least-squares 14-14

15 Interpretation as modified Newton method (notation: x = x (k), x + = x (k+1) ) in Gauss-Newton method (from normal eqs of LS problem on page 14-13) ( m ) 1( m ) x + = r i (x) r i (x) T r i (x)( r i (x) T x r i (x)) = x ( m ) 1( m ) r i (x) r i (x) T r i (x) r i (x) interpretation: take unit step in direction v = H 1 g(x) where H = 2 r i (x) r i (x) T Nonlinear least-squares 14-15

16 compare with the Newton direction at x: where (from page 14-8) v = 2 g(x) 1 g(x) 2 g(x) = 2 ( r i (x) r i (x) T +r i (x) 2 r i (x)) interpretation of GN-method: replace 2 g(x) in Newton s method by H = 2 r i (x) r i (x) T H 2 g(x) if the residuals r i (x) are small advantage: no need to evaluate 2 r i (x) H is always positive semidefinite Nonlinear least-squares 14-16

17 Gauss-Newton method with backtracking given initial x, tolerance ǫ > 0, parameter α (0,1/2) repeat 1 evaluate r i (x) and r i (x) for i = 1,,m, and calculate r := r 1(x) r m (x), A := r 1(x) T r m (x) T 2 if 2A T r ǫ, return x 3 v := (A T A) 1 A T r 4 t := 1 while m r i(x+tv) 2 > r 2 +α(2r T Av)t, t := t/2 5 x := x+tv until maximum number of iterations is exceeded Nonlinear least-squares 14-17

18 example (same problem as page 14-6) x (k) x v x (1) x (0) k u local convergence is slower than Newton s method Nonlinear least-squares 14-18