Journal of Applied Mathematics and Physics, 2013, 1, 68-72
http://dx.doi.org/10.4236/jamp.2013.14013 Published Online October 2013 (http://www.scirp.org/journal/jamp)
Copyright © 2013 SciRes. JAMP
An Efficient Pattern Search Method
Xiaoli Zhang, Qinghua Zhou, Yue Wang
College of Mathematics and Computer Science, Hebei University, Baoding, 071002, Hebei, China
Email: z xl.hbu@gmail.com, qinghua.zhou@gmaill.com
Received June 2013
ABSTRACT
Pattern search algorithms is one of most frequently used methods which were designed to solve the derivative-free op-
timization problems. Such methods get growing need with the development of science, engineering, economy and so on.
Inspired by the idea of Hooke and Jeeves, we introduced an integer
m
in the algorithm which controls the number of
steps of iteration update. W e mean along the descent direction, we allow the algor ithm to go ahead
m
steps at most to
explore whether we can get better solution further. The experiment proved the strategy’s efficiency.
Keywords: Unconstrained Optimization; Derivative-Free Optimization; Pattern Search Methods; Pos itive Bases
1. Introduction
In this paper, we consider the unconstrained minimiza-
tion pro bl e m
min( )
n
xR
fx
where
:
n
fR R
, is continuously differentiable, but
the information about the gradient of
f
is either un-
available or unreliable. There are lots of problems where
derivatives are unavailable but we also want to do some
optimizations. The diversity of applications comes from
different complicated backgrounds with economics, en-
gineering, mathematics, finance, and so on (see [1-3] for
instance).
In such cases, derivative-free optimization methods
(also named direct search methods) which neither com-
pute nor approximate derivatives play an important role.
The reader is referred to see [4-6]. In [5], the author in-
troduced an ingenious idea for a generalized pattern
search method and gave convergence analysis. It in-
cludes several known algorithms as its special cases. Fa-
miliar with the analysis of th e property of the gen eralized
method, the author developed two new classes of pattern
search methods [6].
Inspired by the idea of Hooke and Jeeves [7], we im-
proved the method of [6] by introducing an integer
m
.
We mean, if a step is successful (the value of
f
de-
crease), then the same direction maybe also be proved
successfully at the current point. So, we allow the algo-
rithm to explore the same direction further. On the other
hand, if it always goes ahead along one direction until it
can not improve the value of
f
any more, it likely
neglects additional information which other directions
can offer. To balance these two aspects, we introduce an
integer
m
be used to control iteration steps which we
mean that we allow the algorithm to iterate at most
m
steps along the same direction.
Next, we would like to present some basic concepts
we need.
2. Pattern Search Methods and Positive
Bases
We use
and
,<⋅⋅>
to represent the Euclidean
norm and inner product, respectively. By abuse of nota-
tion, if
A
is a matrix,
aA
means that the vector
a
is a column of
A
. It will also be convenient to assume
that
12
[, ,, ]
r
aa a
represents, not only the matrix with
r
columns, but also, depending on the context, the set
of
r
vectors
. The identity matrix is
denoted by
I
and its
i th
column by
i
e
. Finally,
we write
e
to represent a vector of ones with appropri-
ate size.
2.1. Positive Bases
We present a few basic properties of positive bases be-
ginning from the theory of positive linear dependence
developed by Davis [8]. The positive span of a set of
vectors
12
[, ,, ]
r
vv v
is the convex cone
1
{:,0,1,2,, }
r
nii i
i
vR vavir
α
=
∈ =≥=
The set
12
[, ,, ]
r
vv v
is said to be positively de-
pendent if one of the vectors is in the convex cone posi-
tively spanned by the remaining vectors, i.e., if one of the
vectors is a positively combination of the others; other-
wise the set is positively independent. A positive basis is
X. L. ZHANG ET AL.
Copyright © 2013 SciRes. JAMP
69
a positively independent set whose positive span is
n
R
.
Alternatively, a positiv e basis for
n
R
can be defined as a
set of nonzero vectors of
n
R
whose positive combina-
tions span is
n
R
, but no proper set does. The following
theorem in [8] indicates that a positive spanning set con-
tains at least
1n+
vectors in
n
R
.
Theorem 1 If
12
[, ,, ]
r
vv v
positively spans
n
R
,
then it contains a subset with
1r
elements that
spans
n
R
.
Furthermore, a positive basis can not contain more
than
2n
elements ([8]). Positive basis with
1n+
and
2n
elements are referred to as minimal and maximal
positive basis respectively.
We present now three necessary and sufficient char a c-
terizations for a set of vectors that spans
n
R
or
spans
n
R
positively ([8]).
Theorem 2 Let
12
[, ,, ]
r
vv v
, with
0
i
v
for
all
1, 2,,ir=
, span
n
R
. Then the following are
equivalent: i)
12
[, ,, ]
r
vv v
positively spans for
n
R
.
ii) For every
1, 2,,ir=
,
i
v
is in the convex
cone positively spanned by the remaining
1r
vectors.
iii) There exist real scalars
12
,,,
r
αα α
with
0
i
α
>
,
{1,, }ir
, such that
1
0
r
ii
i
v
α
=
=
. iv) For every
nonzero vector
n
bR
, there exists an index
i
in
{1,, }ir
for with
0
Ti
bv>
.
The following result provides a simple mechanism for
generating different positive bases. The proof can be
found in [6].
Theorem 3 Suppose
12
[, ,, ]
r
vv v
is a positive ba-
sis for
n
R
and
nn
BR
×
is a nonsingular matrix,
then
12
[ ,,,]
r
Bv BvBv
is also a positive basis
for
n
R
.
From above theorems, we can easily deduce the fol-
lowing corollary.
Corollary 1 Let
12
[, ,,]nn
n
B bbbR
×
= ∈
be a
nonsingular matrix, then
1
[, ]
n
i
i
Bb
=
is a positive
basis for
n
R
.
A trivial consequence of this corollary is that
[,]Ie
is a positive basis.
2.2. Pattern Sea rch Methods
Pattern search methods are characterized by the nature of
the generating matrices and the exploratory moves algo-
rithms. These features are discussed more fully in [5] and
[9].
To define a pattern, we need two components, a basis
matrix and a generating matrix.
The basis matrix can be any nonsingular matrix
nn
BR
×
. The generating matrix is a matrix
k
nP
k
CZ
×
, where
1
k
Pn>+
. We partition the ge-
nerating matrix into components
[ ]
, ,0
k kk
CL= Γ
We require that
k
MΓ∈
, where
M
is a finite set
of integral matrices with full row rank. We will see that
k
Γ
must have at least
1n+
columns. The 0 in the last
column of
k
C
is a single column of zeros.
A pattern
k
P
is then def i n ed by the columns of the
matrix
kk
P BC=
. For convenience, we use the parti-
tion of the generating matrix
k
C
to partition
k
P
as
follows:
[, ,0]
k kkk
P BCBBL== Γ
Given
0, >∆∈∆
kk
R
, we define a trial step
i
k
s
to be any vector of the form
i
kk
i
k
BCs∆=
, where
i
k
C
is a to be any vector of the form
i
kk
i
k
BCs∆=
, column
of
k
C
. Note that
i
k
BC
determines the direction of the
step, while
k
serves as a step length parameter.
At the
k th
iteration process, we define a trial
point as any point of the form
i
kk
i
k
sxx +=
, where
k
x
is the current iteration point.
The following algorithms state the pattern search me-
thod for unconstr a i ne d minim i z a tion.
2.3. Algorithm 1 Pattern Search Method
Let
0n
xR
and 00∆> be give n.
For
0,1,2,k=
Compute
( )
k
fx
.
Determine a step
k
s
using an unconstrained explora-
tory moves algorithm.
If
()( )
kk k
fx sfx+<
, then
1k kk
x xs
+
= +
, otherwise
1kk
xx
+
=
.
Update
k
C
and
k
.
2.4. Algorithm 2 Updating
k
Let , and
{ }
01
,,, ,
l
www Z
0
0w<
,
0,1,2,,
i
wi l≥=
.
Let
0
,
w
θτ
=
{ }
,1, 2,,
i
k
il
λτ
∈=
.
If
()( )
kk k
fx sfx+≥
, then
1kk
θ
+
∆=∆
.
If
()( )
kk k
fx sfx+<
, then
1k kk
λ
+
∆=∆
.
3. Our Algorithm and Numerical Results
In [5], the generating matrix has the form
[ ]
,, ,0
kk kk
CM ML= −
for some
nn×
nonsingular
matrix
k
M
. In light of the above discussion, the nature
of
[ ]
,
k kk
MMΓ= −
as a maximal positive basis is now
revealed.
In [6], the author reduced the number of objective
evaluations in the worst case from
n2
to as few as
1+n
. The choice is to make
k
Γ
include
1+n
col-
umns which are just the minimal positive bases.
In this paper, we simply select the relative parameters
as follows:
[ ]
,,
k
BI Ie=Γ=−
with
( )
T
e1,,1,1=
,
1/2, 1
k
θλ
= =
.
Then, we have all we need to state our algorithm now.
X. L. ZHANG ET AL.
Copyright © 2013 SciRes. JAMP
70
Algorithm 3 Modified Pattern Search Method
(1) Start with
0 00
,, ,0,1xf kmn∆==
and
m
.
(2) Check the stopping criteria.
(3) Let
i
kkkkCxx∆+=
+1
and compute
( )
1+k
xf
.
If
()( )
1kk
fx fx
+<
then go to step (4), else go to
step (5).
(4) If
mmn
, then
kkkkxxxx −+=
+++ 111
and
compute
( )
1+
k
xf
. If
()()
kk
xfxf <
+1
, then
1kk
xx
+
=
,
11
,1
kk
xxmn mn
++
== +
go to step (4);
else
1kk+
∆=∆
,
1
,1
kk
xx kk
+
== +
, go to step (5).
(5) If
1in<+
, then
1ii=+
, go to step (3); else set
11
1
,,1, 1
2
k kkk
xxkk i
++
= ∆=∆=+=
, go to step
(2).
In fact, from the above algorithms, we can see that if
we think any successful step as an iteration , the n
B
in our
algorithm should be
I
(identity matrix) or
( )
,
n
Pij
(A matrix which exchanges the
i th
row (or column)
and the
j th
one of the identity matrix). Whenever a
step is found failure, then
B
is set to be
I
again. It is easy
to know that our choices and settings satisfied the condi-
tions in [5,6]. Then, we would like to state the conver-
gence theorem which is also the same as in [5,6].
Theorem 4 Assume that
( )
0
Lx
is compact and
that
f
is continuously differentiable on a neighborhood
of
( )
0
Lx
. Then for the sequence of iterates
k
x
gener-
ated by algorithm 3, we have
( )
k
lim inffx0
k→∞
∇=
.
Proof: The reader is referred to [5,6].
Re mark :
( )
0
Lx
is Level set defined as follows:
()( )( )
{ }
00
,
n
Lxxfxfxx R=≤∈
.
We tested our algorithm on the 18 examples given by
Moré, Garbow and Hillstrom [9]. The 19-th is our testing
problem at the beginning which we used for testing the
effectiveness of the new algorithm. Its definition is:
( )
2
2
2
1
xxxf +=
. We select
m
to equal 0, 1, 2, 3, 5
and 10 respectively. It is easy to know that when
0m=
,
it is just the traditional pattern search method with posi-
tive basis.
The column “P” denotes the number of the problems,
and ”N ” the number of variables. The numerical results
are given by “F” which denotes the number of function
evaluations. And “f” denotes the final function value we
got when
m
= 2. Additionally, the symbols ×” means
that the algorithm terminates because the number of
function evaluation exceeds 500,000. And for the easy
comparing among the results we rearranged the order of
the number of problems. The stopping condition we select is
6
10
k
∆≤
, (1)
which is different with other relative documents.
We select Equation (1) as the stopping criteria just be-
cause it is simple and easy for understanding. It is
thought that if very small step can not lead to decrease in
function value, then the current iteration point maybe
located in a neighborhood of a local minimum. The algo-
rithm is also terminated if the nu mber of function eva lua-
tions exceeds 500,000. And we test it from three kinds of
initial points, say,
0
x
, 10
0
x
and 100
0
x
. The values
of them are suggested in [10]. We will see that our algo-
rithm is robust and performs the best when m = 2. The
results are represented in the following tables in the Ap-
pendix.
4. Acknowledgements
We are very grateful to the referees for their valuable
comme nt s an d suggestio ns.
REFERENCES
[1] A. Ismael, F. Vaz and L. N. Vicente, “A Particle Swarm
Pattern Search Method for Bound Constrained Global
Optimization,” J. Glob. Optim., Vol. 39, No. 2, 2007, pp.
197-219. http://dx.doi.org/10.1007/s10898-007-9133-5
[2] S. Sankaran, C. Audet and A. L. Marsden, “A Method for
Stochastic Constrained Optimization Using Derivative-Free
Surrogate Pattern Search and Collocation,J. Comput.
Phys., Vol. 229, No. 12, 2010, pp. 4664-4682.
http://dx.doi.org/10.1016/j.jcp.2010.03.005
[3] J. Yuan, Z. Liu and Y. Wu, “Discriminative Video Pattern
Search for Efficient Action Detection,” IEEE Trans. on
PAMI., Vol. 33, No. 9, 2011, pp. 1728-1743.
http://dx.doi.org/10.1109/TPAMI.2011.38
[4] R. M. Lewis, V. Torczon and M. W. Trosset,Direct
Search Methods: Then and Now,” J. Comp. Appl. Math.,
Vol. 124, No. 1-2, 2000, pp. 191-207.
http://dx.doi.org/10.1016/S0377-0427(00)00423-4
[5] V. Torczon, “On the Convergence of Pattern Search Al-
gorithms,” SIAM J. Optim, Vol. 7, 1997, pp. 1-25.
http://dx.doi.org/10.1137/S1052623493250780
[6] R. M. Lewis and V. Torczon, “Rank Ordering and Posi-
tive Bases in Pattern Search Algorithms,” Technical re-
port TR96-71, ICASE, 2000.
[7] R. Hooke and T. A. Jeeves, “’Direct Search’ Solution of
Numerical and Statistica l Problems,” J. ACM., Vol. 8, No.
2, 1961, pp. 212-229.
http://dx.doi.org/10.1145/321062.321069
[8] C. Davis, “Theory of Positive Linear Dependence,” Amer.
J. Math., Vol. 76, 1954, pp. 733-746.
http://dx.doi.org/10.2307/2372648
[9] V. Torczon, “Multi-Directional Search: A Direct Search
Algorithm for Parallel Machines,” PH.D Thesis, Depart-
ment of Mathematical sciences, Rice University, Houston,
TX, 1989.
[10] J. J. Mor, B. S. Garbow and K. E. Hillstrom, “Testing
Unconstrained Optimization Software,” ACM Transac-
tions on Mathematical Software, Vol. 7, 1981, pp. 17-41.
http://dx.doi.org/10.1145/355934.355936
X. L. ZHANG ET AL.
Copyright © 2013 SciRes. JAMP
71
Appendix
Table 1. The results for initial point
0
x
.
P N m = 0, F m = 1, F m = 2, F m = 3, F m = 5, F m = 10,F m = 2, f
2 6
×
×
×
×
×
×
4 2
×
×
×
×
×
×
10 2
×
×
×
×
×
×
12 3
×
130,552 127,201 105,761 119,244
×
8.52906D-2
1 3 5590 70,774 7741 7750 7750 7750 4.34729D-8
3 3 7182 53,656 7412 7413 7413 7413 1.16348D-8
5 3 2825 839 485 2441 2419 2398 534.653
6 3 2651 121 2607 2639 2607 2607 1.20174D-9
7 3 1812 5140 895 1644 1644 1644 0.47141
8 3 176,941 109 110 110 110 110 1.68911D-5
9 3 1326 105 161 1284 1284 1284 4.75148D-2
11 4 14,559 92,655 13,991 14011 14,010 14,010 85822.2
13 3 709 651 665 670 670 670 2.57368D-3
14 3 13,858 27,322 20,993 28,958 95,780 18,656 4.32517
15 3 8310 11,383 8403 8403 8403 8403 2.39103
16 2 9658 20,884 4011 5058 4511 4511 8.91800D-4
17 4 58,246
×
50,618 50,841 50,876 50,876 5.087505D-12
18 4 374 861 423 425 425 425 4.36513D-12
19 2 156 157 157 157 157 157 3.63798D-12
Table 2. The results for initial point 10
0
x
.
P N m = 0, F m = 1, F m = 2, F m = 3, F m = 5, F m = 10, F m = 2, f
2 6
×
×
×
×
×
×
4 2
×
×
×
×
×
×
10 2
×
×
×
×
×
×
14 3
×
395,663 405,203
×
×
×
11.85160
1 3 5599 1494 7750 7759 7759 5194 4.34729D-8
3 3 13,989
×
962 27,048 10,557 9172 0.28108
5 3 87,860 410 376 362 343 329 84.988
6 3 2676 86 86 86 86 86 32.4938
7 3 1812 5165 918 1664 1661 1658 0.47140
8 3 177,085 255 10301 1711 8963 240,840 1.65986D-5
9 3 1453 679 1605 1622 1569 1522 3.3929D-6
11 4 18,145 112,591 16,484 788 16337 16276 85822.2
12 3 19,716 5833 14,437 32,520 32520 32,520 8.7120D-4
13 3 717 708 762 767 767 767 2.5736D-3
15 3 8571 11,765 8571 8637 8570 8539 2.39102
16 2 9685 20,904 4026 5074 4525 4523 8.91800D-4
17 4 98,178 78,527 85,858 86,334 86,388 86,383 5.087505D-7
18 4 473 507 494 494 495 495 6.13756D-11
19 2 201 190 188 189 190 190 3.63798D-12
X. L. ZHANG ET AL.
Copyright © 2013 SciRes. JAMP
72
Table 3. The results for initial point 100
0
x
.
P N m = 0, F m = 1, F m = 2, F m = 3, F m = 5, F m = 10, F m = 2, f
5 3
×
101 98 95 95 92 10222
10 2
×
×
×
×
×
×
14 3
×
×
×
×
×
×
1 3 5689 1584 7840 7849 7849 7849 403472D-8
2 6 142,287 66,879 59,418 58,982 51,958 51,426 8.70496
3 3 856 1167 393 352 463 430 0.41085
4 2 134 152 152 152 152 152 1.000
6 3 4375 35,350 1405 9375 3863 5833 91.756
7 3 2212 5390 1098 1819 1796 1775 0.47140
8 3 178,525 1245 11,067 2470 9622 241,459 1.6598D-5
9 3 1633 789 1695 1700 1638 1579 3.392D-6
11 4 50,455 162,471 61,685 32,205 29,500 27,933 85822.2
12 3 81 81 81 81 81 81 32.835
13 3 608 1042 1091 1095 1095 1095 2.5736D-3
15 3 11181 29,597 9944 10,002 9706 10,387 2.391
16 2 9955 21,084 4176 5208 4645 4628 8.91800D-4
17 4 133,549 88,233 115,142 108,730 115,674 115,623 5.087505D-7
18 4 1013 869 977 768 868 724 6.13756D-11
19 2 651 505 458 434 415 394 3.63798D-12