Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013 1497<br />
T 1 T<br />
m<strong>in</strong> ∇ f( x)<br />
d + d Hd<br />
2<br />
T<br />
s.. t g ( x) +∇ g ( x) d = 0, j∈A⊆<br />
I,<br />
j<br />
j<br />
where the so-called work<strong>in</strong>g set A ⊆ I is suitably<br />
determ<strong>in</strong>ed. If d 0<br />
= 0 and λ ≥ 0 ( λ is said to be the<br />
correspond<strong>in</strong>g KKT multiplier vector.), the algorithm<br />
stops. The most advantage of these algorithms is merely<br />
necessary to solve QP sub-problems with only equality<br />
constra<strong>in</strong>ts. However, if d<br />
0<br />
= 0 , but λ < 0 , the algorithm<br />
will not implement successfully. In [10], proposed an<br />
SQP method for general constra<strong>in</strong>ed optimization. Firstly,<br />
make use of the technique which handle the general<br />
constra<strong>in</strong>ed optimization as an <strong>in</strong>equality parametric<br />
programm<strong>in</strong>g, then, consider a new quadratic<br />
programm<strong>in</strong>g with only equality constra<strong>in</strong>ts as follow:<br />
T 1 T<br />
m<strong>in</strong> ∇ f( x)<br />
d + d Hd<br />
2<br />
T<br />
s.. t g ( x) +∇ g ( x) d =−m<strong>in</strong>{0, π ( x)}, j∈J( x).<br />
j<br />
j<br />
Where π ( x)<br />
is a suitable vector, J ( x ) is a suitable<br />
approximate active set. But the QP problems may no<br />
solution under some conditions. Recently, Zhu [14]<br />
Consider the follow<strong>in</strong>g QP sub-problem:<br />
T 1 T<br />
m<strong>in</strong> ∇ f( x)<br />
d + d Hd<br />
2<br />
T<br />
s.. t p ( x) +∇ g ( x) d = 0, j∈L.<br />
j<br />
j<br />
where p ( x ) is a suitable vector, L is a suitable<br />
j<br />
approximate active, which guarantees to hold that if<br />
d<br />
0<br />
= 0 , then x is a KKT po<strong>in</strong>t of (1), i.e., if d<br />
0<br />
= 0 , then<br />
it holds that λ ≥ 0 . Depended strictly on the strict<br />
complementarity, which is rather strong and difficult for<br />
test<strong>in</strong>g, the superl<strong>in</strong>ear convergence properties of the<br />
SQP algorithm are obta<strong>in</strong>ed. For avoid<strong>in</strong>g the superl<strong>in</strong>ear<br />
convergence depend strictly on the strict complementarity,<br />
Another some SQP algorithms (see [15]) have been<br />
proposed, however it is regretful that these algorithms are<br />
<strong>in</strong>feasible SQP type and nonmonotone. In [16], a feasible<br />
SQP algorithm is proposed. Us<strong>in</strong>g generalized projection<br />
technique, the superl<strong>in</strong>ear convergence properties are still<br />
obta<strong>in</strong>ed under weaker conditions without the strict<br />
complementarity.<br />
We will develop an improved feasible SQP method for<br />
solv<strong>in</strong>g optimization problems based on the one <strong>in</strong> [14].<br />
The traditional FSQP algorithms, <strong>in</strong> order to prevent<br />
iterates from leav<strong>in</strong>g the feasible set, and avoid Maratos<br />
effect, it needs to solve two or three QP sub-problems<br />
like (2). In our algorithm, per s<strong>in</strong>gle iteration, it is only<br />
necessary to solve an equality constra<strong>in</strong>ed quadratic<br />
programm<strong>in</strong>g, which is very similar to (4). Obviously, it<br />
is simpler to solve the equality constra<strong>in</strong>ed QP problem<br />
than to solve the QP problem with <strong>in</strong>equality constra<strong>in</strong>ts.<br />
In order to void the Maratos effect, comb<strong>in</strong>ed the<br />
generalized projection technique, a height-order<br />
correction direction is computed by an explicit formula,<br />
and it plays an important role <strong>in</strong> avoid<strong>in</strong>g the strict<br />
(3)<br />
(4)<br />
complementarity. Furthermore, its global and superl<strong>in</strong>ear<br />
convergence rate is obta<strong>in</strong>ed under some suitable<br />
conditions.<br />
This paper is organized as follows: In Section II, we<br />
state the algorithm; the well-def<strong>in</strong>ed of our approach is<br />
also discussed, the accountability of which allows us to<br />
present global convergence guarantees under common<br />
conditions <strong>in</strong> Section III, while <strong>in</strong> Section IV we deal<br />
with superl<strong>in</strong>ear convergence. F<strong>in</strong>ally, <strong>in</strong> Section V,<br />
numerical experiments are implemented.<br />
II. DESCRIPTION OF ALGORITHM<br />
The active constra<strong>in</strong>ts set of (1) is denoted as follows:<br />
I( x) = { j∈ I | g ( x) = 0, j∈ I}.<br />
(5)<br />
Now, the follow<strong>in</strong>g algorithm is proposed for solv<strong>in</strong>g<br />
the problem (1).<br />
Algorithm A:<br />
Step 0 Initialization:<br />
Given a start<strong>in</strong>g po<strong>in</strong>t<br />
0<br />
x ∈ X , and an <strong>in</strong>itial<br />
n n<br />
symmetric positive def<strong>in</strong>ite matrix H 0<br />
∈ R × . Choose<br />
1<br />
parameters ε0<br />
∈(0,1), α∈(0, ), τ ∈ (2,3) . Set k = 0 ;<br />
2<br />
Step 1. Computation of an approximate active set J<br />
k<br />
.<br />
Step 1.1. For the current po<strong>in</strong>t<br />
0<br />
j<br />
k<br />
x<br />
k<br />
ε ( x ) = ε ∈ (0,1).<br />
i<br />
∈ X , set i = 0,<br />
Step 1.2. If det( A ( x k ) T A( x k )) ≥ ε ( x<br />
k ) , let<br />
i i i<br />
k k k<br />
Jk<br />
= J( x ), Ak<br />
= A( x ), i( x ) = i , and go to Step 2.<br />
Otherwise go to Step 1.3, where<br />
k k k<br />
J ( x ) = { j∈I | −ε<br />
( x ) ≤ g ( x ) ≤0},<br />
i i j<br />
k k k<br />
A( x ) = ( ∇g ( x ), j∈J ( x )).<br />
i i i<br />
k 1 k<br />
Step 1.3. Let i = i+ 1, εi( x ) = εi<br />
− 1( x ) , and go to<br />
2<br />
Step1. 2.<br />
k<br />
Step 2. Computation of the vector d<br />
0<br />
.<br />
Step 2.1<br />
B A A A v v j J B f x<br />
T −1<br />
T k k k<br />
k<br />
= (<br />
k k) k<br />
, = (<br />
j, ∈<br />
k) = −<br />
k∇<br />
( ),<br />
k k<br />
⎧ , 0<br />
k ⎪ − vj<br />
vj<br />
<<br />
k k<br />
pj = ⎨<br />
p = ( pj, j∈Jk).<br />
k k<br />
⎪⎩ g<br />
j( x ), vj<br />
≥ 0<br />
Step 2.2 Solve the follow<strong>in</strong>g equality constra<strong>in</strong>ed QP<br />
k<br />
Sub problem at x :<br />
k T 1 T<br />
m<strong>in</strong> ∇ f( x ) d + d Hkd<br />
2<br />
k k T<br />
s. t. p +∇ g ( x ) d = 0, j∈J<br />
.<br />
j j k<br />
k<br />
Let d0<br />
be the KKT po<strong>in</strong>t of (8), and<br />
k k<br />
b = ( b , j∈J<br />
) be the correspond<strong>in</strong>g multiplier vector.<br />
j<br />
k<br />
k<br />
If d<br />
0<br />
= 0 , STOP. Otherwise, CONTINUE;<br />
(6)<br />
(7)<br />
(8)<br />
© 2013 ACADEMY PUBLISHER