You are on page 1of 5

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 7, JULY 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
1

A Geometric Programming Approach for The


Automotive Production in Turkey
Ersoy ÖZ and İbrahim GÜNEY

Abstract—Geometric Programming is an important class of nonlinear programming problems. And it is characterized by


objective and constraints that have a special form. There may be a nonlinear objective function and nonlinear constraints in a
Geometric Programming model. Nonlinear optimization models have been solved by different methods. Among them,
Geometric Programming is an efficient method for solving a particular type of nonlinear problems. Thus it may be advantageous
over linear programming. In this work, Cobb-Douglas production function is considered as a Geometric Programming model.
The 2009 production data of three automotive firms in Turkey are used in the model. The objective of the work is to estimate the
number of employees that will minimize the production costs, and the capital input, using Geometric Programming solution
method. For the estimations, ggplab toolbox that uses the interior-point algorithm and works under the Matlab software is used.

Index Terms—Constrained Optimization, Convex Programming, Nonlinear Programming, Optimization.

——————————  ——————————

1 INTRODUCTION

G EOMETRIC Program (GP) is a class of nonlinear


optimization with many useful theoretical and com-
putational properties. Although GP in standard
hausted by the early 1980s, the perid of 1980-1998 was
relatively quiet. Over the last few years, however, GP
started to receive renewed attention from the operations
form is apparently a non-convex optimization problem, it research community. In particular, we now have very
can be readily turned into a convex optimization prob- efficient algorithms to solve GP, either by general pur-
lem, hence a local optimum is also a global optimum, the pose convex optimization solvers or by more specialized
duality gap is zero under mild conditions, and a global methods [12].
optimum can be computed very efficiently. Convexity Production functions models can usually be expressed
and duality properties of GP are well understood, and and solved as GP models. The initial significant studies
large-scale, robust numerical solvers for GP are available. on production functions are as follows: [11], [3], [8], [4],
Furthermore, special structures in GP and its Lagrange [9], [1]. The Cobb-Douglas model which is taken as the
dual problem lead to computational acceleration, distri- basis of this study is presented in [6].
buted algorithms, and physical interpretations [13]. Two of the best known production functions utilized
In 1961, Zener published a seminal paper [7] and ob- by economists are the so-called Leontief and Cobb-
served that some engineering design problems can be Douglas functions. These functions are named after the
formulated as optimization of ‘generalized polynomials’, individuals mainly responsible for their popularization in
and that if the number of terms exceed the number of the economics literature. From a historical perspective,
variables by one, the optimal design can be found by they are interesting because they have been widely used
solving a system of linear equations. In 1967, Duffin, Pe- in both theoretical and empirical production analyses and
terson, and Zener published the book [17] that started the other fields of economics. However, their widespread use
field of GP as a branch of nonlinear optimization. Several in theoretical analysis only came about after their exten-
important developments of GP took place in the late sive use in empirical analysis. As such, it represents an
1960s and 1970s. GP was tied with convex optimization excellent example of applied production analysts crystal-
and Lagrange duality, and was extended to include more lizing a means of research for the entire economics pro-
general formulations beyond posynomials. Several nu- fession [16].
merical algorithms to solve GP were proposed and tested. The Turkish automotive industry was first established
And researchers in mechanical engineering, civil engi- in the 1950s and the development of the sector occured
neering, and chemical engineering found successful ap- when the factories built in the late 1960s and the early
plications of GP to their problems. 1970s have reached a certain capacity and proficiency
However, as researches felt that most of the theoretical, rate. The first production of automotive industry
algorithmic and application aspects of GP had been ex- products in Turkey was in 1954, made by Türk Willys
Overland Ltd. The first serious production of automobiles
———————————————— started in 1966 with the Anadol automobile. Later on, the
 Ersoy ÖZ, Department of Technical Programs, Yildiz Technical Universi- firms TOFAŞ and OYAK-RENAULT were respectively
ty, Turkey. effective in the development and strengthening of the
 İbrahim GÜNEY, Department of Statistics, Istanbul Aydin University,
Turkey. automotive industry in Turkey. Today, 15 main firms are
active in the sector and they significantly help the country
to develop with around 25.000 employment that they
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 7, JULY 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 2

provide [24]. under a minute, on a small desktop computer [20].


Interior-point methods for GPs are very robust. They
require essentially no algorithm parameter tuning, and
2 GEOMETRIC PROGRAMMING
they require no starting point or initial guess of the op-
2.1 Monomial and Posynomial Functions timal solution. They always find the (true, globally) op-
Let x1,..., xn denote n real positive variables, and timal solution, and when the problem is infeasible (i.e.,
x  ( x1,..., xn ) a vector with components xi . A real valued the constraints are mutually inconsistent), they provide a
function f of x , with the form certificate showing that no feasible point exists. General
f ( x )  cx1a1 x2a2 ... xnan (1) methods for nonlinear programming (NLP) can be fast,
where c  0 and ai   , is called a monomial function. but are not guaranteed to find the true, global solution, or
Any positive constant is a monomial, as is any variable. even a feasible solution when the problem is feasible. An
Monomial functions are closed under multiplication and initial guess must be provided, and can greatly affect the
division. A monomial function raised to any power is also solution found, as well as the solution time. In addition,
a monomial function. algorithm parameters in general purpose NLP solvers
The term ‘monomial’ differs from the standard defini- have to be carefully chosen.
tion of ‘monomial’ used in algebra. In algebra a monomial The main trick to solving a GP efficiently is to convert
has the form (1), but the exponents ai must be nonnega- it to a nonlinear but convex optimization problem, i.e., a
tive integers, and the coefficient c is one. In this paper problem with convex objective and inequality constraint
the coefficient can be any positive number, and the expo- functions, and linear equality constraints. Efficient solu-
nents can be any real numbers, including negative and tion methods for general convex optimization problems
fractional. are well developed [20].
A sum of one or more monomials, i.e., a function of the For the model that is addressed in this study, ggplab
form K toolbox that uses the interior-point algorithm is used.
f ( x )   ck x1a1k x2a2 k ... xnank (2) This toolbox works under the Matlab software [19].
k 1

where ck  0 , is called a posynomial function [21].


3 THE MODEL
2.2 Geometric Program In much economic research dealing with issues of growth
A GP is an optimization problem of the form and productivity, econometric analysis is done based on a
production function. The most commonly used produc-
minimize f0 ( x ) tion function is the Cobb-Douglas. This incorporate capi-
subject to fi ( x )  1, i  1,..., m, (3) tal and labor as the factor inputs which are usually consi-
g i ( x )  1, i  1,..., p, dered the major resources in the production process. Oc-
casionally, capital and labor are supplemented by land,
where fi are posynomial functions, g i are monomials, materials or energy to attempt to complete the production
and xi are optimization variables. The problem (3) is a function [18].
geometric program in standard form. In a standard form The firm uses n inputs (labor, coal, iron, ore, etc.) to
of GP, the objective must be posynomial function and it produce a single output, let x be the column vector of
must be minimized; the equality constraints can only inputs,
have the form of a monomial function equal to one, and x   ( x1,..., xn )
the inequality constraints can only have the form of posy- and let q be the output. The firm’s production function is
nomial function less than or equal to one. represented by
GP, a mathematical programming technique, provides q  f ( x )  f ( x1,..., xn ) (4)
a powerful tool for solving nonlinear problems where giving output as a function of its inputs. (4) assumes
nonlinear relations can be well represented by exponen- nothing but the existence of a maximum output corres-
tial or power functions. Its attractive structural properties ponding to any combination of inputs.
as well as its elegant theoretical basis have led to a num- Let r be a row vector of given prices of the inputs,
ber of interesting engineering applications and the devel- r  (r1, r2 ,..., rn )
opment of numerous useful results. If the primal problem and p the given price of the output.
is in a standard form of GP, then a minimizing solution to A firm is in competitive situation if it can buy and sell
that problem can be obtained by solving the dual max- in any quantities at exogenously given prices, which are
imization. This is because there is a strong duality theo- independent of its production decisions. In other words,
rem for GP problems [22]. the competitive firm is a price taker [14].
The firm benaves so as to maximize the profit  , given
2.3 Solution of Geometric Program as the difference between revenue, pq , and cost, given as
The main motivation for GP modeling is the great effi- the total expenditure on all inputs,
ciency with which optimization problems of this special n
rx   r j x j .
form can be solved. To give a rough idea of the current j 1
state of the art, standard interior-point algorithms can The problem of the (competitive) firm can then be
solve a GP with 1000 variables and 10000 constraints in stated as the following mathematical programming prob-
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 7, JULY 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 3

lem: TABLE 1
maximize x  ( x )  pf ( x )  rx VARIABLES AND PARAMETERS OF THE MODEL
subject to x  0.
Another version of this problem often used in produc-
tion theory is based on the assumption of the given out-
*
put level q . The firm is trying to minimize its cost, M , of
*
the inputs used to produce q . The expenditure of the
firm is given by
M  r1x1  r2 x2  ...  rn xn
and the mathematical programming problem is then
minimize x M ( x )
subject to f ( x )  q * and x  0. (5)

Closely interrelated to this formulation is the problem


of the firm with a given level of expenditure or prespeci-
*
fied budget, M , and an objective function that maximiz-
es the production q . Thus the firm chooses levels of in-
puts so as to maximize output, subject to a budget con-
straint: 4 DATA
maximize x f ( x ) The data that will be used in the application belong to the
(6) automotive companies that were producing in Turkey in
subject to r1x1  r2 x2  ...  rn xn  M * and x  0.
2009. From these companies, three of them that made the
For illustration of (5) and (6), we consider the technol- most sale and that spent the most for crude material and
ogy defined by the Cobb-Douglas production function labor force are chosen. These companies are Ford Otosan,
(with just two inputs): Oyak Renault and Tofaş and they will be shown by F1, F2
q  f ( x1, x2 )  ax1 x2 and F3 respectively.
with a  0,0    1, and 0    1. The data of price of labor input, output level,
In this case, (5) leads to the following mathematical pro- employment of labor input and cost of inputs used to
gramming model: produce q * are taken from [2]. The capital input values
minimize x1, x2 r1x1  r2 x2 are taken from [10] for F1, from [15] for F2 and from [23]
(7) for F3. Capital input values are the issued capitals of the
subject to ax1 x2  q * and x1  0, x2  0. respective firms. Price of capital inputs values are
*
It is to find, for a particular output level q and with a weighted avarege interest rates for deposits and corres-
given structure of input prices, what input levels would pond to the price of capital. This data is taken from [5].
constitute the cheapest way of producing this output and TABLE 2
what would be the minimum cost. This question can be INPUT DATA FOR THE MODEL
answered for all possible levels of output and the mini-
mum cost would be depend on the level of output to be
produced.
Assuming that the technology is defined by the Cobb-
Douglas production function, the solution of (7) will yield
the cost function C(r , q ) , expressing minimum cost as a
function of input prices r and output level q . A slight
modification of the constraint in (7) leads to the following
geometric programming problem:
minimize x1, x2 M ( x )  r1x1  r2 x2
q *    TL: Turkish Lira
subject to x1 x2  1 (8) The notations in Table 1 are the explanat
a
x1  0, x2  0. In economy, any produced production tool is capital
which is a factor that carries out production. The capital
The notations in Table 1 are the explanations of the pa- factor helps people during the production process and
rameters and variables in (8). increases efficiency. The capital, as a production factor,
covers physical capital goods. Along with this, the
present monetary capital is not involved in physical
production but is used to obtain the phsyical capital.
The activities that an enterprice carries out to renew or
increase its equipment are called “investment”. The
payments made to organizations or individuals in
exchange of the provided finance to accomplish these
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 7, JULY 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 4

activities are called ‘interest’. This shows that the interest TABLE 4
is the price of capital. THE COMPARISON OF THE ESTIMATED AND ACTUAL VALUES

5 APPLICATION
GP is a nonlinear programming problem which might
have nonlinear objective and constraint functions. The
model analysed in this study is Cobb-Douglas model
which is the basis of production functions. In economics,
the Cobb-Douglas functional form of production func-
tions is widely used to represent the relationship of an
output to inputs. For this model the objective function is a
posynomial function which contains two monomial 6 RESULTS AND CONCLUSION
terms. The first constraint function in (8) is:
The estimated x1 values for all of the three firms is 3.70
q *    because of the following constraint in the model:
x1 x2  1
a x1  Automotive Industry Employment
and the second constraint function is:
If this constraint had not been added to the model, the
x1  Automotive Industry Employment.
x1 values would have been much below 3.70. As a result,
The second constraint is used for the purpose of model’s
realistic results would not have been obtained. The esti-
consistency. The data to be used in the model might have
mated x2 values are very close to the actual x2 values.
9-10 digits, therefore all the data is used by their loga-
However estimated M values are very different than the
rithms. The variables and parameters in the model set up
actual M values.
as shown below:
The Cobb-Douglas production function that is ana-
minimize x1, x2 M ( x )  r1x1  r2 x2 lysed in this study, gives consistent results except for M (
q *    the cost of inputs used to produce q * ) values, in the light
subject to x1 x2  1
a (9) of the existing data. The reason of the inconsistency of the
x1  Automotive Industry Employment M values is that there are too many parameters which
explain the objective value and that these parameters are
x1  0, x2  0 not existent in the model. The parameters that determine
the value of the objective function are r1 , r2 and the va-
are r1 value which refers to the logarithm of the price of
riables are x1 , x2 . However, the number of the parame-
labor input, r2 value which refers to the logarithm of the
ters and variables for producing q * are too many. Thus, it
price of capital input and q * value which refers to the
logarithm of the output levels. The parameter of a (Total is obvious that there would be more consistent results if
Factor Productivity) in constraint is taken as 1. Besides the objective and constraint functions could be revised as
the value of the automotive industry employment is de- including more parameters and values.
The successful results that are obtained in this study by
termined as of a minimum of 5000.
GP solving method of the Cobb-Douglas production func-
The values of  ,  , x1 , x2 and M are estimated for
tion will contribute to further studies in this realm.
F1, F2 and F3 by using the data and the model in (9). As
mentioned above, ggplab toolbox that works under the
Matlab software is used for calculating algorithms. The REFERENCES
results obtained are in Table 3. [1] A.A. Alchian and H. Demsetz, “Production, Imformation Costs, and
Economic Organization,” The American Economic Review, vol. 62, pp.
TABLE 3 777-795, 1972.
ESTIMATED VALUES [2] Automotive Manufacturers Association, “General and Statistic-
al Information Bulletin of Automotive Manufacturers 2010-I,”
http://www.osd.org.tr/cata2010.pdf. (2010). (29.05.2010)
[3] A. Zellner, J. Kmenta, and J. Dreze, “Specification and Estimation of
Cobb-Douglas Production Function Models,” Econometrica, vol. 34,
pp. 784-795, 1966.
[4] A. Zellner and N.S. Revankar, “Generalized Production Functions,”
The Review of Economic Studies, vol. 36, pp. 241-250, 1969.
[5] Central Bank of the Republic of Turkey, “Weighted Average
Interest Rates For Deposits,” http://www.tcmb.
The comparison of x1 , x2 and M estimated values gov.tr/yeni/bgm/yfaagrmev/agrmev_TRL.html. (2010).
and the logarithms of the actual values are in Table 4. (14.05.2010)
[6] C.W. Cobb and P.H. Douglas, “A Theory of Production,” American
Economic Association, vol. 18, no. 1, pp. 139-165.
[7] C. Zener, “A mathematical aid in optimizing engineering de-
sign,” Proceedings of National Academy of Sciences, vol. 47, pp.
537-539, 1961.
[8] D.J. Aigner and S.F. Chu, “On Estimating the Industry Production
Function,” The American Economic Review, vol. 58 , pp. 826-839, 1968.
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 7, JULY 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 5
[9] D.J. Hodges, “A Note on Estimation of Cobb-Douglas and CES Pro-
duction Function Models,” Econometrica, vol. 37, pp. 721-725, 1969.
[10] Ford Otosan, “Corporate Governance Reports,”
http://www.fordotosan.com.tr/downloads/yatirimciiliskileri/
2009_Yili_Faaliyet_Raporu.pdf. (2010). (14.05.2010)
[11] H.B. Chenery, “Engieerin Production Functions,” The MIT Press, vol.
63, pp. 507-531, 1949.
[12] K.O. Kortanek, X. Xu, and Y. Ye, “An infeasible interior-point
algorithm for solving primal and dual geometric programs,”
Mathematical Programming, vol. 76, pp. 155-181, 1996.
[13] M. Chiang, Geometric Programming for Communication Systems.
Hanover: Now Publishers Inc., pp.1-132, 2005.
[14] M. Luptacik, Mathematical Optimization and Economic Analysis.
London: Springer Science+Business Media, pp. 7-9, 2000.
[15] Oyak Renault, “Annual Report,” http://www.oyak.com.tr
/OyakWEBTR/faaliyet_raporlari/20100419011748.Faaliyet_Ra
poru_2009.pdf. (2010). (14.05.2010)
[16] R.G. Chambers, Applied Production Analysis: A Dual Approach.
New York: Cambridge University Press, pp. 15-16, 1988.
[17] R.J. Duffin, E.L. Peterson, and C. Zener, Geometric Programming:
Theory and Applications. New York: Wiley, 1967.
[18] R.N. Mefford, “Introducing Mangement Into The Production
Function,” The Review of Economics and Statistics, vol. 68, no. 1,
pp. 96-104, Feb. 1986.
[19] S. Boyd, “GGPLAB: A Simple Matlab Toolbox for Geometric
Programming,” http://www.stanford.edu/~boyd /ggplab/.
2006. (15.03.2010)
[20] S. Boyd and L. Vandenberghe, Convex Optimization. United
Kingdom: Cambridge University Press, pp. 561-614, 2004.
[21] S. Boyd, S.J. Kim, L. Vandenberghe and A. Hassibi, “A Tutorial
on Geometric Programming,” Optim. Eng., vol. 8, pp. 67-127,
2007.
[22] S.T. Liu, “A geometric programming approach to profit max-
imization,” Applied Mathematics And Computation, vol. 182, pp.
1093-1097.
[23] Tofaş Türk Otomobil Fabrikası A.Ş., “Annual and Interim Re-
ports,” http://tofas.com.tr/backup/Documents/tr/pdf/ TO-
FASFAALIYETRAPORU_2009.pdf. (2010). (14.05.2010)
[24] T.R. Prime Ministry State Planning Organization, “Türkiye’de
Otomotiv Sanayii Gelişme Perspektifi,” www.dpt.
Gov.tr/DocObjects/Download/3310/gelisme.pdf. (2002).
(19.04.2010)

Ersoy ÖZ has obtained his Ph.D. degree in Operations Research at


Marmara University in Turkey in 2009. His works are about markov
chains, hidden markov models and geometric programming. He is
lecturer in the Computer Programming department at Yildiz Technic-
al University.

İbrahim GÜNEY has obtained his Ph.D. degree in Applied Mathe-


matics at Uludag University in Turkey in 1991, associate professor
degree in Statistics in 2002 and professor degree in 2008. His works
are about statistical shape analysis, correspondence analysis, statis-
tical manifolds and geometric programming. He is professor in the
Statistics department at Istanbul Aydin University.

You might also like