Global Optimization of Mixed Integer Signomial Fra
Global Optimization of Mixed Integer Signomial Fra
problems journals.sagepub.com/home/mac
Abstract
The main issue in the present article is to investigate how to solve a mixed integer fractional signomial geometric pro-
graming problem (MIFSGP). In the first step to achieving this idea, we must convert a fractional signomial programing
problem into a non-fractional problem via a simple conversion technique. Then, a convex relaxation with a new modified
piecewise linear approximation with integer break points as a pre-solve method is used to reach an integer global opti-
mum solution. A few numerical examples are included to illustrate the advantages of the proposed method
Keywords
Global optimization, geometric programing, signomial fractional programing, mixed integer programing, and piecewise
linear approximation.
Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
1212 Measurement and Control 57(8)
P Q a
problems globally with as small as distance from the solution denoted as CðxÞ = Jj= 1 cj ni= 1 xj ji where cj
solution of the original problem. In a valuable work by and aji2 Rnf0g.
Westerlund,14 a few helpful transformation techniques The variables in a signomial function must be
are considered in deterministic global optimization for Positive.
signomial programing problems. In addition, he applied This study considers a MISFP problem with the fol-
the most effortless power transformations to convexify lowing form:
posynomial and signomial geometric terms. Although
the GGPECP algorithm used by Westerlund in 2003 PðxÞ
Minimize fðxÞ =
was used to solve the convexified and underestimated QðxÞ
problems to reach a globally optimal solution, in his
method, he used too many constraints in expressing a S:t Ax4b
piecewise linear function, which may cause heavy com-
putational burdens. Therefore we have presented an Bk ðxÞ + Ck ðxÞ40, k = 1, 2, . . . , K
efficient technique by combining power transforma-
tions tool to convexify the non-convex signomial terms X = ðx1 , x2 , . . . , xn Þ, 04xi 4xi 4
xi
and a more accurate modified linearization technique
which can give mixed integer solution as close as possi-
ble to the original problem. By applying our propose, Where P(x) and Q(x) represent twice–differentiable
we observed not only less number of iterations but also signomial functions and Q(x) . 0.
got the integer solution as well. The matrix A and vector b are constants, and Bk
There are two ways to convex a positive signomial and Ck represent convex and non–convex signomial
term (posynomial) with power transformation, so- functions, respectively. xi and xi indicate the lower
called Negative power transformation (NPT) and posi- and upper bounds of a real and bounded variable xi.
tive power transformation (PPT). Since the transformation approach is valid for signo-
Although using NPT is more accessible than PPT, mial constraints, therefore the signomial fractional
the PPT technique makes a convex form of signomial expression QPððxxÞÞ in objective function can be rewritten in
function tighter than NPT to original non-convex func- the constraint form as below:
tions, as explained in Westerlund et al.14 Tsai (2005)
proposed a new approach to solve a general signomial Minimize xn + 1
fractional programing problem to reach a global opti-
mum. He used NPT to convexify posynomial terms and PðxÞ
S:t xn + 1 40, QðxÞ . 0 8x
superior piecewise linearization techniques. We now use QðxÞ
the PPT technique to convexify the non-convex posyno-
mial term along with piecewise linear approximations. S:t Ax4b
In this work, we have applied a modified model for
piecewise linear functions of one and two variables Bk ðxÞ + Ck ðxÞ40, k = 1, 2, . . . , K
which perform significantly better than the standard
binary models that Vilma and Nemhauser15 proved the
x = ðx1 , x2 , . . . , xn Þ, 04xi 4xi 4xi , i = 1, . . . , n
efficiency of this superior piecewise linearization tech-
nique. Therefore, this work discusses a mixed integer
global optimal solution to convexify every non-convex The problem is further converted to:
signomial term. Univariate transformations are applied
PðxÞ xn + 1 :Q(x)40
in such a way to convexify and underestimate with this
piecewise linear approximation of the inverse transfor-
mation of f(x), where x lies in the interval [a0, am] such Where xn + 1 is a new real and bounded variable.
that a0, am2 Z with m + 1 integer breakpoints. In this (2.2) is a signomial function and may be non–
article, applying integer breakpoints for integer vari- convex.
ables is studied in detail as a pre-solution method to
reach the integer solution for those variables.
This issue develops a global optimization method Convexification strategies
with fewer iteration numbers and CPU time than other The problem of convexification a non-convex signo-
methods to obtain a mixed integer solution in solving a mial term, such as inverse and exponential transforma-
Signomial Fractional Programing problem (SFP). tions, is discussed in detail in Shen and Yu2,7 Lundell
et al.16, Chang1 In this study, we use a particular case
Problem formulation of the inverse transformation technique suggested by
Westerlund. Convexity requirements for negative and
A mixed integer signomial fractional programing prob- positive signomial terms are given in the following
lem (MISFP) seeks an approximate global optimum propositions.9
Nejad et al. 1213
(iii) Yn
ð2Þ
i=k+1 i
X bPi
Now, concerning k, there are two modes: Where L(Xi) is a piecewise linearization approximation of
1
And L(Xi) is a piecewise linearization approximation of And LðXi Þ5Xi For i = 2, . . . , k ð9Þ
1
Xi = xib i = k + 1, . . . , n ð4Þ Since the first transformation variable in the left-hand
inequality of (iii) is an increasing function where the
next k21 variables in the transformed signomial term
Hint: Without losing generality, the signomial terms have negative powers, it corresponds to decreasing
can be rewritten by (i), inverse transformations (3), and functions. Hence, it was found that (8) and (9) are valid
(4) for a, b . 0. Now, with assumptions (ii) and (iii), if the inverse transformation (6) is provided a concave
we want to prove that Z can be convexified and function and the inverse transformations (7) are offered
underestimated. by convex functions too. Thus, (iii) is fully satisfied
Proof. Since all powers in (iii) are positive with parameter a is chosen, such as the a51 for (6) to be a
a, b . 0, the inequality (iii) will be valid if concave function and b . 0 for (7) to be a convex
function.
LðXi Þ5Xi i = 1, . . . , n ð5Þ Since a51 and b . 0, in the expression (ii),
P P
ap1 b ki= 2 pi + ni= k + 1 pi 51
We find that (5) and thus (iii) will be valid if (3) and (4) The first term is positive and the second and third
are convex functions; it is fulfilled if a is chosen in are negative. It is also valid for any sufficiently large a
0 \ a41 and b . 0, so they are satisfied by conditions and small b.
of (ii) and remark 1. Hence, according to conditions (ii) and (iii), z is a
Thus, Z in (i) is a convex term referring to proposi- convex w.r.t.
tion 3.
Piecewise linear approximation
Theorem 1: A positive signomial term (posynomial)
z = cxp11 xp22 . . . xpnn . Where pi 50 (for i = 1, . . . , k), A custom way to Solve piecewise linear approximation
is to apply the special order sets (SOS). However, this
study used a much better approach in fractional pro-
Pi \ 0 (for i = k + 1, . . . , n), and p1 5p2 5 . . . 5pn graming, which is more efficient than ordinary meth-
can be convexified and underestimated as follows: ods. Therefore, this work uses a tight relaxation
method of approximating piecewise linear functions.15
1 a p1 Q 1 b !pi Remark17: An injective function B : f1, 2, . . . , mg
k Q
i) z = c x1a : xib : ni= k + 1 xpi i . ! f0, 1gu , u = log2 m, where the vectors B(p) and
i=2
B(p + 1) differ in at most one component for all
P
k Pn
ii) ap1 b pi + pi 51, where a51, b . 0. p 2 f1, 2, . . . , m 1g.
i=2
Q
i=k+1
Q Let B(p) = ðu1 , u2 , . . . , uu Þ, 8uh 2 f0, 1g, k = 1, 2, . . . u.
iii) cðLðX1 ÞÞap1 : ki= 2 (LðXi Þ)bPi : ni= k + 1 xpi i 4cXap
1 :
1
and B(0) = B(1).
Qk Qn
i=2 Xbp
i
i
: i=k+1 xpi i S+ (k) = fp j 8B(p) and B(p + 1), uk = 1, p = 1, 2,
. . . , m 1g [ fp j 8B(p), ui = 1, p 2 f0, mgg
Where L(Xi) represents piecewise linear approximations S (k) = fp j 8B(p) and B(p + 1), uY = 0, p = 1, 2,
of . . . , m 1g [ fp j 8B(p), uk = 0, p 2 f0, mgg.
1
X1 = xa1 ð6Þ Theorem 217: The Given a univariate function f(x),
a0 4x4am , denote L(f(x)) as the piecewise linear func-
tion of f(x), where
and
1 a0 \ a1 \ a2 \ \ am be the m + 1 integer break
Xi = xib , for i = 2, . . . , k: ð7Þ points of L(f(x)) can be expressed as
X
m X
m X
m
L(f(x)) = f ap wp , x = ap w p , wp = 1,
Proof. Without losing generality, z can be rewritten by p=0 p=0 p=0
1 1
b
(i) by choosing X1 = x1 , Xi = xi , i = 2, . . . , k a,
a
X X
Qk bpi Qn
b . 0, so we have z = cXap 1
1 : i = 2 Xi : i = k + 1 xpi i . wp 4uk , wp 41 uk , 8wp 2 R+ ,
Inequality (iii) should be valid if p2S+ (k) p2S (k)
Since we are looking for a global integer solution for a Let a1 = 0:5, a2 = 0:5, a3 = 2:4 so it needs to
class of mixed integer nonlinear programing problems, power transformations as: X1 = x2 1 , X2 = x1
2
0:5 0:5 1:2 0:5 0:5 2:4
we use in this study integer breakpoints in the piecewise and X3 = x3 therefore x1 x2 x3 = X1 X2 X3
linearization technique to approximate the inverse To convexify the non-convex negative signomial
transformation functions for every integer decision term 2 x2 x3 x4 by referring to proposition 1, we
variable. This strategy can solve the mixed integer con- have:
vexified signomial programing problems faster than the a(1 + 1 + 1)41 with 04a41 so we have a4 13,
custom (SOS). Thus, the integer results were achieved and the inverse transformation variables yi = x3i and
for integer variables with no greater dominations. for the non-convex signomial term 2x2 x3 x4 yields to
1 1 1
2Y32 Y33 Y34 , and for another signomial term:
44
Illustrative examples x2:1 1:3 10
2 x3 x4 We have: 4:4a41 ) a4 34 , Zi = ðxi Þ
10
21 13 10
so x2:1 1:3
2 x3 x4 is convexified as Z2 Z3 Z4 .
44 44 44
Example 1. Consider the following non-convex MISFP The original MISFP problem is reformulated as
problem. The tolerable error is specified as e = 0.02 follows:
from Chang1: Minimize x4
1 1 1 21 13 10
S:t X^0:5
1 X^0:5
2 X^2:4 ^3 ^3 ^3 ^44 ^44 ^44
3 2Y2 Y3 Y4 Z2 Z3 Z4 40
2 1
x1 x0:5
2 x3
1:2
X^1 = L x1 , X^2 = L x2 , X^3 = L x0:5
Min 44
3
44
2x2 x3 + x2 x1:3
2:1
3 3
ð10Þ ^
Yi = L xi , i = 2, 3, 4, Z4 = L x4 , Z^3 = L x133 ,
^ 10
S:t 44
2x1 + 3x2 + x3 58 Z^2 = L x21 2
ð1Þ (5)
3x1 + x2 55 ð11Þ x1 , x4 2 R, x2 , x3 2 Z
x1 (0.2,1.4,2.6,3.8,5)
X
4 X
4 x2 (1,2,3,4,5)
x2 = aP wP + 5 , Y^2 = (aP )0:5 wP + 5 , x3 (1,2,3,4,5)
P=0 P=0 x4 (0.01,0.49,0.97,1.45,1.91)
X
4
Z^1 = (aP )4 wP , aP 2 ½0:05, 2
P=0
Nejad et al. 1217
Table 2. The comparison of results for the shock absorber design problem.
Variable Arora (1989)18 Coello (1999)19 Ray and Saini (2001)20 J-F Tsai (2005)9 Proposed method