quasi-newton methods of optimization lecture 2. general algorithm n a baseline scenario algorithm u...
TRANSCRIPT
Quasi-Newton Methods of Optimization
Lecture 2
General Algorithm
A Baseline Scenario• Algorithm U (Model algorithm for n-
dimensional unconstrained minimization). Let xk be the current estimate of x*.
– U1. [Test for convergence] If the conditions for convergence are satisfied, the algorithm terminates with xk as the solution.
– U2. [Compute a search direction] Compute a non-zero n-vector pk, the direction of the search.
General Algorithm
– U3. [Compute a step length] Compute a scalar ak, the step length, for which f(xk + akpk )<f(xk).
– U4. [Update the estimate of the minimum] Set xk+1 = xk + ak pk, k=k+1, and go back to step U1.
• Given the steps to the prototype algorithm, I want to develop a sample problem that we can compare the various algorithms against.
4321
1.4
4.3
3.2
2.1
32100
max
xxxxst
xxxxx
General Algorithm
– Using Newton-Raphson, the optimal point for this problem is found in 10 iterations using 1.23 seconds on the DEC Alpha.
Derivation of the Quasi-Newton Algorithm An Overview of Newton and Quasi-Newton
Algorithms• The Newton-Raphson methodology can be
used in U2 in the prototype algorithm. Specifically, the search direction can be determined by:
p f x f xk xx k x k 2 1
( ) ( )
Derivation of the Quasi-Newton Algorithm
• Quasi-Newton algorithms involve an approximation to the Hessian matrix. For example, we could replace the Hessian matrix with the negative of the identity matrix for the maximization problem. In this case the search direction would be:
p I n f xk x k ( ) ( )
Derivation of the Quasi-Newton Algorithm
• This replacement is referred to as the steepest descent method. In our sample problem, this methodology requires 990 iterations and 29.28 seconds on the DEC Alpha.
– The steepest descent method requires more overall iterations. In this example, the steepest descent method requires 99 times as many iterations as the Newton-Raphson method.
Derivation of the Quasi-Newton Algorithm
– Typically, the time spent on each iteration is reduced. Again, in the current comparison each the steepest descent method requires .123 seconds per iteration while Newton-Raphson requires .030 seconds per iteration.
Derivation of the Quasi-Newton Algorithm
• Obviously substituting the identity matrix uses no real information from the Hessian matrix. An alternative to this drastic reduction would be to systematically derive a matrix Hk which uses curvature information akin to the Hessian matrix. The projection could then be derived as:
p H f xk k x k 1 ( )
Derivation of the Quasi-Newton Algorithm Conjugate Gradient Methods
• One class of Quasi-Newton methods are the conjugate gradient methods which “build” up information on the Hessian matrix.
– From our standard starting point, we take a Taylor series expansion around the point xk + sk
x k k x k xx k kf x s f x f x s( ) ( ) ( )2
Derivation of the Quasi-Newton Algorithm
for some sk N(xk,). Solving this expression for the term involving the Hessian yields
x k k x k xx k k
k x k k x k k xx k k
f x s f x f x s
s f x s f x s f x s
( ) ( ) ( )
' ( ( ) ( )) ' ( )
2
2
Derivation of the Quasi-Newton Algorithm
y f x s f xk x k k x k ( ) ( )
y f x sk xx k k2 ( )
y B sk k k 1
Derivation of the Quasi-Newton Algorithm One way to generate Bk+1 is to start with the
current Bk and add new information on the current solution
B B uv
y B uv sk k
k k k
1 '
( ' )
Derivation of the Quasi-Newton Algorithm
u v s y B sk k k k( ' )
uv s
y B sk
k k k 1
'( )
B Bv s
y B s vk kk
k k k 1
1
'( ) '
Derivation of the Quasi-Newton Algorithm
• The Rank-One update then involves choosing v to be yk + Bksk. Among other things, this update will yield a symmetric Hessian matrix:
B By B s s
y B s y B sk kk k k k
k k k k k k
1
1
( )'( )( )'
Derivation of the Quasi-Newton Algorithm
• Other than the Rank-One update, no simple vector will result in a symmetric Hessian. An alternative is to reconfigure the Hessian by letting the numeric be the 1/2 the sum of a numeric approximation plus itself transposed. This procedure yields the general update:
B Bv s
y B s v v y B sy B s s
v sv vk k
kk k k k k k
k k k k
k
1 2
1
'( ) ' ( )'
( )'
( ' )'
DFP and BFGS
• Two prominent conjugate gradient methods are the Davidon-Fletcher-Powell (DFP) update and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update.
– In the DFP update v is set equal to yk yielding
B Bs B s
B s s By s
y y s B s w w
wy s
ys B s
B s
k kk k k
k k k kk k
k k k k k k k
kk k
kk k k
k k
1
1 1
1 1
''
'' ' '
' '
DFP and BFGS
– The BFGS update is then
B Bs B s
B s s By s
y yk kk k k
k k k kk k
k k 1
1 1
''
''
DFP and BFGS
A Numerical Example• Using the previously specified problem and
starting with an identity matrix as the original Hessian matrix, each algorithm was used to maximize the utility function.
Bt
1 0 0
1 0
1
Bt*
. . .
. .
.
5275 2885 0718
6085 0954
2244
DFP and BFGS
• In discussing the difference in step, I will focus on two attributes.
– The first attribute is the relative length of the step (the 2-norm).
– The second attribute is the direction of the step. Dividing each vector by its 2-norm yields yields a normalized direction of the search
st . . .7337 9766 2428
st* . . . 4 3559 4 3476 4 3226
DFP and BFGS
x1 x2 x3 x1 x2 x3
Newton-Raphson 4.36 4.35 4.32 7.52 0.58 0.58 0.57Conjugate Gradient 0.73 0.98 0.24 1.25 0.59 0.78 0.19
Relative Performance
– The Rank One Approximation• Iteration 1
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6690 3842 1536
5540 1783
9287
2780 1276 0511
2501 0592
2280
x1 x2 x3 x1 x2 x3
Newton-Raphson 6.75 7.63 4.83 11.27 0.60 0.68 0.43Conjugate Gradient 4.32 5.01 2.00 6.91 0.62 0.73 0.29
Relative Performance
• Iteration 2
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6315 4256 1776
5083 2048
9134
0628 0171 0106
0612 0113
0846
x1 x2 x3 x1 x2 x3
Newton-Raphson 10.85 11.69 6.08 17.07 0.64 0.68 0.36Conjugate Gradient 7.73 8.91 3.79 12.39 0.62 0.72 0.31
Relative Performance
– PSB• Iteration 1 B
B
t
t
. . .
. .
.
. . .
. .
.
*
6703 3860 1504
5565 1827
9365
2780 1276 0511
2501 0592
2280
x1 x2 x3 x1 x2 x3
Newton-Raphson 10.85 11.69 6.08 17.07 0.64 0.68 0.36Conjugate Gradient 7.73 8.91 3.79 12.39 0.62 0.72 0.31
Relative Performance
• Iteration 2
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6328 4274 1745
5109 2096
9223
0629 0171 0106
0612 0114
0850
x1 x2 x3 x1 x2 x3
Newton-Raphson 10.85 11.70 6.07 17.07 0.64 0.69 0.36Conjugate Gradient 7.72 8.91 3.78 12.38 0.62 0.72 0.30
Relative Performance
– DFP• Iteration 1 B
B
t
t
. . .
. .
.
. . .
. .
.
*
7187 4517 0326
6455 3424
12232
2780 1276 0511
2501 0592
2280
x1 x2 x3 x1 x2 x3
Newton-Raphson 6.75 7.63 4.83 11.27 0.60 0.68 0.43Conjugate Gradient 4.14 5.01 1.76 6.73 0.61 0.74 0.26
Relative Performance
• Iteration 2
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6788 4945 0589
6021 3766
12194
0653 0177 0119
0602 0124
0971
x1 x2 x3 x1 x2 x3
Newton-Raphson 6.75 7.63 4.83 11.27 0.60 0.68 0.43Conjugate Gradient 4.14 5.01 1.76 6.73 0.61 0.74 0.26
Relative Performance
– BFGS• Iteration 1
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6771 3952 1338
5690 2051
9768
2780 1276 0511
2501 0593
2280
x1 x2 x3 x1 x2 x3
Newton-Raphson 10.81 11.93 5.80 17.11 0.63 0.70 0.34Conjugate Gradient 7.52 9.06 3.40 12.26 0.61 0.74 0.28
Relative Performance
• Iteration 2
B
B
t
t
. . .
. .
.
. . .
. .
.
*
6391 4369 1585
5238 2333
9644
0634 0172 0109
0610 0115
0871
x1 x2 x3 x1 x2 x3
Newton-Raphson 10.84 11.74 6.02 17.08 0.63 0.69 0.35Conjugate Gradient 7.69 8.94 3.71 12.36 0.62 0.72 0.30