25.01.2015 Views

Download Full Issue in PDF - Academy Publisher

Download Full Issue in PDF - Academy Publisher

Download Full Issue in PDF - Academy Publisher

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

1498 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />

Step 3<br />

Computation of the feasible direction with descent<br />

Where<br />

δ<br />

ek<br />

k<br />

d :<br />

k k T −1<br />

d = d0 − δ<br />

kAk( Ak Ak)<br />

ek<br />

(9)<br />

T | Jk<br />

|<br />

= (1, ,1) ∈R<br />

, and<br />

|| d || ( d ) H d<br />

k k T k<br />

0 0 k 0 k T −1<br />

T k<br />

k<br />

= , μ =−( Ak Ak) Ak∇f( x )<br />

kT k<br />

2 | μ ek<br />

||| d0<br />

|| + 1<br />

Step 4. Computation of the high-order revised direction<br />

k<br />

d :<br />

where<br />

d A A A d e g x d (10)<br />

k T −1<br />

k k k<br />

=− ( ) (||<br />

0<br />

|| τ<br />

k k k k<br />

+<br />

J<br />

( + )),<br />

k<br />

k k k k k k T k<br />

g<br />

( x + d ) = g ( x + d ) −g ( x ) −∇g ( x ) d .<br />

J J J<br />

k k k J k<br />

Step 5. L<strong>in</strong>e search:<br />

Compute t k<br />

, the first number t <strong>in</strong> the sequence<br />

1 1 1<br />

{1, , , ,...}<br />

satisfy<strong>in</strong>g<br />

2 4 8<br />

2<br />

f ( x k + td k + t d k ) ≤ f( x k ) + αt∇f( x k ) T d<br />

k , (11)<br />

g x td t d j I<br />

(12)<br />

k k 2 k<br />

j<br />

( + + ) ≤0, ∈ .<br />

Step 6. Update:<br />

Obta<strong>in</strong> H<br />

k +<br />

by updat<strong>in</strong>g the positive def<strong>in</strong>ite matrix<br />

1<br />

H k<br />

us<strong>in</strong>g some quasi-Newton formulas. Set<br />

k 1 k k 2 k<br />

x + = x + t d + t d , and k = k+ 1 . Go back to step 1.<br />

k<br />

Throughout this paper, follow<strong>in</strong>g basic assumptions<br />

are assumed.<br />

H2.1 The feasible set X ≠Φ, and functions f ( x ),<br />

g<br />

j<br />

( x),<br />

j∈ I are twice cont<strong>in</strong>uously differentiable.<br />

H2.2 ∀ x ∈ X , the vectors { ∇g ( x), j∈I( x)}<br />

are<br />

l<strong>in</strong>early <strong>in</strong>dependent.<br />

Lemma 2.1 Suppose that H2.1and H2.2 hold, then<br />

1) For any iteration, there is no <strong>in</strong>f<strong>in</strong>ite cycle <strong>in</strong> step 1.<br />

2) If a sequence { x k } of po<strong>in</strong>ts has an accumulation<br />

po<strong>in</strong>t, then there exists a constant _ ε > 0<br />

ε<br />

kik ,<br />

_<br />

> ε for k large enough.<br />

j<br />

such that<br />

Proof.<br />

1) Suppose that the desired conclusion is false, that is<br />

to say, there exists some k, such that there is an <strong>in</strong>f<strong>in</strong>ite<br />

cycle <strong>in</strong> Step 1, then we obta<strong>in</strong>, ∀ i = 1, 2, , that Aki<br />

,<br />

is<br />

not of full rank, i.e., it holds that<br />

det( A A ) = 0, i = 1,2, , (13)<br />

T<br />

ki , ki ,<br />

And by (6), we can know that Jki<br />

, + 1<br />

⊆ Jki<br />

,<br />

. S<strong>in</strong>ce there<br />

are only f<strong>in</strong>itely many choices for J<br />

ki ,<br />

⊆ I , it is sure that<br />

~<br />

J ≡ J L for i large enough. From (6) and (13),<br />

ki , + 1 ki , k<br />

with i →∞, we obta<strong>in</strong><br />

~<br />

L k<br />

k = I ( x ), det( A A ) = 0.<br />

T<br />

k k<br />

I( x ) I( x )<br />

This is a contradiction to H 2.2, which shows that the<br />

statement is true.<br />

2) Suppose K is an <strong>in</strong>f<strong>in</strong>ite <strong>in</strong>dex set such that<br />

*<br />

{ x } → x . We suppose that the conclusion is false,<br />

k k∈K<br />

i.e., there exists<br />

for<br />

~<br />

L<br />

k<br />

Let<br />

~<br />

'<br />

'<br />

K ⊆ K K<br />

(| | =∞ ) , such that<br />

ε<br />

ki<br />

→ k∈K k →∞<br />

'<br />

,<br />

0, , .<br />

k<br />

Lk = Jk, i k −1. From the def<strong>in</strong>ition of ε<br />

ki , k<br />

, it holds,<br />

k∈<br />

K ' ,<br />

k large enough, that<br />

~<br />

T T k<br />

~ ~ ε<br />

ki , k j k<br />

Lk<br />

Lk<br />

det( A A ) = 0, −2 ≤ g ( x ) ≤0, j∈ L . (14)<br />

S<strong>in</strong>ce there are only f<strong>in</strong>itely many choices for sets<br />

⊆ I , it is sure that there exists<br />

such that<br />

~ ~<br />

''<br />

k<br />

, ( )<br />

'' ' ''<br />

K ⊆ K (| K | =∞ ) ,<br />

L ≡ L k∈ K , for k large enough.<br />

Denote ~ *<br />

A = { ∇g ( x )| j∈ L<br />

~<br />

} , then, let<br />

from (14), it holds that<br />

j<br />

~ T ~ ~<br />

* *<br />

=<br />

j<br />

= ∈ ⊆<br />

k∈K '' , k →∞ ,<br />

det( A A) 0, g ( x ) 0, j L I( x ).<br />

This is a contradiction to H 2.2, too, which shows that<br />

the statement is true.<br />

Lemma 2.2 For the QP sub-problem (8) at x<br />

k , if<br />

k<br />

d<br />

0<br />

= 0 , then x k<br />

k<br />

is a KKT po<strong>in</strong>t of (1). If d0 ≠ 0 , then<br />

k<br />

x computed <strong>in</strong> step 4 is a feasible direction with descent<br />

k<br />

of (1) at x .<br />

Proof.<br />

By the KKT conditions of QP sub-problem (8), we<br />

have<br />

k k k<br />

∇ f( x ) + Hkd0<br />

+ Akb<br />

= 0,<br />

k k T k<br />

p +∇ g ( x ) d = 0, j∈J<br />

,<br />

j j 0<br />

k<br />

k<br />

If d<br />

0<br />

= 0 , we obta<strong>in</strong><br />

k k k<br />

∇ f ( x ) + A b = 0, p = 0, j∈<br />

J ,<br />

k j k<br />

k<br />

Thereby, from (7) and x ∈ X,<br />

∀ k implies that<br />

k<br />

k<br />

g ( x ) = 0, v ≥0, j∈<br />

J .<br />

j j k<br />

b k B k k<br />

k<br />

f x v<br />

In addition, we have =− ∇ ( ) = , <strong>in</strong> a word,<br />

we obta<strong>in</strong><br />

k k<br />

∇ f ( x ) + A b<br />

k k<br />

= 0, g ( x ) = 0, b ≥0, j∈<br />

J ,<br />

k j j k<br />

k<br />

Let bj<br />

= 0, j∈ I \ Jk<br />

, which shows that x k is a KKT<br />

po<strong>in</strong>t of (1).<br />

k<br />

If d0 ≠ 0 , we have<br />

k T k T k k<br />

g ( x ) d = A d = −p − δ e ,<br />

So,<br />

Jk<br />

k k k<br />

k T k k T k kT k<br />

∇ f ( x ) d = − ( d ) H d + b p ,<br />

0 0 k 0<br />

© 2013 ACADEMY PUBLISHER

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!