Utah State University

DigitalCommons@USU

All Graduate Plan B and other Reports Graduate Studies

1973

Integer Programming by Cutting Planes Methods

Sung-Yen Wu
Utah State University

Follow this and additional works at: https://digitalcommons.usu.edu/gradreports

Part of the Applied Statistics Commons

Recommended Citation
Wu, Sung-Yen, "Integer Programming by Cutting Planes Methods" (1973). All Graduate Plan B and other
Reports. 1161.
https://digitalcommons.usu.edu/gradreports/1161

This Report is brought to you for free and open access by

the Graduate Studies at DigitalCommons@USU. It has
been accepted for inclusion in All Graduate Plan B and
other Reports by an authorized administrator of
DigitalCommons@USU. For more information, please
contact [email protected].
 INTEGER PROGRAMMING l3Y CUTTING

PLANES METHODS

by

Sung- Yen Wu

A report submitted in partial fulfillment

of the requirements for the degree

of

MASTER OF SCIENCE

1n

Applied Statistics

Plan B

Approved:

UTAH STA TE UNIVERSITY

Logan, Utah

1973
 ii

AC I<NOW I ,EDG MEN TS

I wish to express my sincere thanks to Dr. Bartell C. Jensen

for his valuable suggestions and guidance in the preparation of this

report. I would also like to thank Professor Ronald V. Canfield and

Dr. Grace Y. Chen for their time spent and support given in this

e ffort.

A great deal of appreciation and thanks go to my parents for

th e ir support in my graduate studies at Utah State University.

Sung-Y'en Wu
 iii

TABLE OF CONTENTS

Page

ACKNOWLEDGMENTS ii

LIST OF TABLES . iv

V
LIST OF FIGURES .

INTRODUCTION . . 1

BASI C C ONCEPT OF LINEAR PROGRAMMING • 5

Description ...... . . . 5
Slack, surplus and artificial variables 6
The simplex method 7
The dual simplex algorithm . . . . 13

INT E G ER PROGRAMMING BY CUTTING PLANES

METOODS ..... . 14

Description 14
Formulation methods and solutions • 15

NUMERICAL EXAMPLES 23

COMPARISON OF THE EFFECIENCY, USAGE, AND

RELATIONSHIP OF DIFFERENT CUTTING PLANES
METHODS 34

Summary 35

REFERENCES 37

APPENDIX .. 38

Computer Programs . 39
 lV

LIST OF TABLES

Table Page

1. The initial primal feasible tableau 23

2. The optimal linear programming tableau with no cut 24

3. The optimal tableau after the first Gomory's cut 25

4. Computation of D! 27
3
5. Computation of D . 28
5
6. Computation of D~ . 28

7. The optimal tableau after the first Kianfer's cut 29

8a. Gomory's algorithm 36

8b. Kianfar's algorithm 36

 V

LIST OF FIGURES

Fi gure Page

l. Cutting Planes Methods on two dimensions . . . 4

 INTRODUCTION

Linear programming 1s a relatively new, very important

branch of modern mathematics and is about twenty five years old.

In this day and age, most planners and decision makers will

acknowledge that some linear optimization problems are worth the

expense and trouble to solve. Using linear programming technique

as a tool to make decision plannes are able to greatly redice cost

or increase profit for any project under consideration.

Since Dr. George B. Dantzig published his first paper on the

simplex method in 1947, progress in that field has been rapid. Al-

though the first applications were military in nature, it was not long

b e fore it became apparent that there were important economic and

industrial applications as well.

Typical areas of application rn linear programming are inven-

tory control, quality control, production scheduling, forecasting,

maintenance and repairs, process design, accounting procedures,

capital budgeting, the diet problem, machine- loading problem, and

the transportation problem.

Integer linear programming was first introduced by Dantzig,

Fulkerson, and Johnson in 1954 in their work on solving the travel-

ing salesman problem and subsequently by Markowitz and Manne in

 2

their work on the solution of discrete programming problems. In

1958 Gomory developed a systematic procedure for obtaining cuts

which can be shown to lead to an optimal solution in a finite number

of steps, when all of the variables are constrained to be integers.

Later, in 1960, the algorithm for solving the mixed-integer case

was also develop ed .

In many practical problems, some or all the variables are

restricted to integer values. Integer programming was introduced

to solve such kind of problems. Integer programming can be used

to solve such problems as the fixed-charge problem, discrete pro-

gramming problems, project planning, manpower scheduling, and

the traveling salesman type problems as well as capital budgeting

problems. The most interesting integer programming problems are

those with variables taking only one of the values O or 1.

At present, many algorithms are available for solving linear

programs of the type discussed here. Four of the most famous

methods are:

1. Cutting-plane algorithms (Gomory, Balas, Kianfar)

2. Group theory for the all integer problem (Gomory, Shapiro)

3. Decomposition (Benders)

4. Enumerative

a. Implicit enumeration for the pure integer problem

(Balas)
 3

b. Branch and Bound (Land and Doig, Litter, Dakin,

Tomlin).

The intention of this study is to review some of tlw cutting

plane methods and to compare one with another in terms of the nUITiber

of interations required and the computation time which results from

each method.

There are two methods that will be considered. The first is

Gomory's method which adds a new constraint, one at a time, to the

original constraint system to yield a sequence of new linear program-

ming problems until the optimal solution has satisfied the integerality

rc•qui rcment (Hadley, l 964 ).

The second method is Kianfar's stronger inequalities for 0, 1

integer programming using Knapsack Functions (Kianfar, 1971 ).

In Gomory's method, the integer points outside the 0, 1 space

can limit the parallel movement of the hyperplane of the cut toward

the solution set, furthermore it does not allow this hyperplane to

rotate for a deeper cut. This method removes these two limitations

to form a stronger c ut in solving 0, 1 integer programming problem.

The cut means to add a new constraint to the original linear program-

ming problem such that some areas in the convex set are cut off.

(See Figure 1)

This report surveys the literature on basic linear programming,

the simplex method and dual simplex algorithm used in integer

 4

-- . -·-····
....- ----- X
I
-; ):,l(J_ry I _, L, l L,

Figure 1. Cutting Planes Methods on two dimensions

programming. It also illustrates two methods of finding solutions to

0, 1 integer linear programming problem by cutting planes methods.

Finally, computer programs will be written for the two methods

to compare their efficiency one with the other. Efficiency as defined

h e r e is the time required on an electronic computer or the number of

it e rations and arithmetic operations. Moreover, an efficient algorithm

may be one which requires relatively little computer core storage.

 5

BASIC CONCEPT OF LINEAR PROGRAMMING

Description

The goal in solving a linear programming problem is to

identify an optimal solution and its associated value of the objective

function.

Mathematically speaking, we have m inequalities or equations

rn r variables (m and r are independent) of the form:

max. or min. z = . + CrXr (1 )

subject to the constraints

b., i = 1, ... ,m ( 2)
l

where the a .. , c. and b. are assumed known constants, and (1) is

lJ J l
c alled the objective function. Usually, some or all of the variables

are restricted to be non-negative. Sometimes, it may be required

that some or all of the variables are allowed to take only integral

values .

When ( 1) and ( 2) are used as a model for a decision problem,

the variables (x., i = 1, . , r) represent the decision maker's

1

possible policy choices, the relations (2) ex press the constraints,

and the objective function is the measure of effectiveness.

 6

Any set of x. which satisfies the set of constraints (2) is called

J
a solution. Any solution which satisfies the non-negativity restrictions

is called a feasible solution. Any feasible solution which maximizes

the z of (2) is called an optimal feasible solution.

The problem its elf may be infeasible, if no solution can satisfy

every co nstraint. Another possibility is that the problem has no finite

maximum value of objective function Z. Such a problem is unbounded

or has an unbounded optimal solution ( Hadley, 1962).

Slack, surplus and artificial variables

We note that an inequality of the type

r
:E ah .x . < bh ( 3)
j= 1 J J -

can be converted into an equation by adding a slack variable x h'

r+
where x h > 0.
r+ -

r
/21ah{j + xr+h = bh ( 4)

Next consi der inequalities having a > sign

r
:E ak.X. > bk ( 5)
J J -
j=l

Equation (5) can be rearranged to

 7

r
~ ak.x - x = bk (6)
j=l J J r+k

by adding a surplus variable x k' where x k > 0.

r+ r+ -
The effect of this change is that the inequality constraints

with~ or > signs have been converted into equations, so that a

system of simultaneous linear equations is obtained. The adjusted

problem has the same set of optimal solutions as the original one

(Hadley, 1962).

Usually it starts rn initial basic solution by setting the basis

matrix B = I.

If some constraints do not require the addition of a slack

variable, the initial basic feasible solution may not be easy to find.

In such cases, artificial variables and vectors are added to

the set of constraints to ensure an initial basic feasible solution to

a linear programming problem can be found.

The simplex method

In solving an integer linear programming problem by the use

of cutting plans. the simplex method seems more suitable and common

than the revised simplex method. The reason is in the simplex method,

it trans forms all the y., xB, z. - c. and z at each iteration.

J J J
-1
Where y. = B a., J = 1, ... , n, xB = B-lb
J J

z = c y the basis matrix B All such

j B j,
 8

information in the simplex method tableau is needed to form a cut in

integer programming.

Although the revised simplex method has the advantage of speed

and storage saving over the standard simplex method, it does not yield

all the information to form a cut. This is due to the fact that only the

basis inverse is transformed at each iteration.

The simplex method is a scheme which systematically evaluates

the extreme points of the convex set in terms of the objective function

until a terminal status is reached.

Three types of terminal status can be identified.

(a) a finite optimal solution; that is an optimal basic feasible

solution.

(b) an infinite optimal solution; that is an unbounded solution.

(c) the problem has no feasible solution; that is the constraints

can not simultaneously be satisifed.

The simplex method starts with a basis that yields a basic

feasible solution. It then moves on to a new basis associated with a

better basic feasible solution which increases the value of the maximi-

zing function (or vis vasa) by means of replacing one of the current

basis vectors with a non- basis vector. The non- basic vector to enter

the basis will yield the greatest contribution to the improvement of

the objective function if it is a maximizing function. On the other hand,

the current basis vector to be removed is selected according to a rule

which insures the continued feasibility of the new solution.

 9

After the vectors to enter and leave the basis have been found,

we start to form a new basis by using transformation formulas. Once

th(' new basis is ~stablish~d, it forms the new starting point for re-

peating the same evaluative process to determine a still better basic

feasible solution. This continues until the final stage is reached.

Computational procedure for the simplex mothod (Hadley, 1962).

Step 1. One wishes to solve the linear programming problem

r
~ a .. x. b., i=l, ... ,m (7)
j=l lJ J 1

to maximize the linear objective function

r
Z = 2_; C . X , (8)
j= 1 J J

with the non-negativity restrictions

X. > 0, (j= 1, .. , , r)
J -

For the convenience of computation, at the beginning check

to see that all the b are non-negative. If any of the b. are

i 1

negative multiply it by -1 to obtain b.> O. Next add slack or

1-

surplus variables to the inequality constraints

Then

r
~ a .. x . < b.
j= 1 lJ J- 1

will bcome
 10
r
~ a . .X . +X . = b ., X . > 0 ( 9)
j= l lJ J r+1 1 r+1 -

and

r
~ a .. X . > b.
j= l lJ J - 1

will become

r
~a .. x . -x . = b., X . > 0 ( 1 0)
j=l lJ J r+1 1 r+1 -

every sla c k or surplus variable is given a price zero. After

the conversion the problem will become

Ax = b, x > 0, max z = ex

where

Step 2. Check to see whether the matrix A contains an m x m identity

matrix, assumes r(A) = m. If the identity matrix does not

exist, add the necessary number of artificial variables and

artificial vectors to obtain an identity matrix. The two phase

method is used to treat artificial vectors. The phase I drive

all artificial variables to zero and the phase II maximize the

actual objective function z.

In phase I, assign prices of zero to the legitimate variables

 11

and of -1 to the artificial variables. Now maximize the

function

s
z = ~ (- 1 )x . =- 1X (1 1 )
i=l at a

where X is an s-component column vector containing the

a

artificial variables. Notice, that z < 0 is always true,

~::
and the n1aximmn value of z is zero.
,:{
At the end of phase I, we may have max z < 0. This

means the problem has no feasible solution. If we have

,,,
max z -·-= 0 at the end of phase I, we have a basic feasible

solution to the original problem, and all the artificial

variables have been driven to zero.

In phase II, we assign its actual price c. to each legitimate

J
variable and a price of zero to any artificial variables which

may appear in the basis at a zero level. The function to be

optimized is the actual objective function z.

Step 3. Construct an initial tableau, and compute z - c from

j j

C a. - C ( l 2)
B J j

Step 4. If all z. - c. > 0, there are two possibilities. If some of the

J J-

xB in the basis do not satisfy the non-negativity restrictions

we have an infeasible s elution, otherwise an optimal basic

feasible solution is obtained.

 12

If one or more (z. - c.) < 0, proceed by determining the

J J
vector ak in A but not in B to enter the basis from

= min ( z. - c . ) , z. - c . < 0 ( 13)

J J J J J

After ak is selected , check yik" if y.k < 0 for all i, there

l -

have an unbounded solution.

St e p 5. If yik > 0 for at least one i, proceed by determining the

vector to be removed from the basis. The criteria that

insures feasibility is

X X 'l
Br i Bi '}
= min I
\ yik > 0 I ( 14)
yrk i yik
\

wh e re the subscript y identifies the vector to leave the basis.

St e p 6 . Us e the v e ctor transformation formuli to find the new basic

11
feasible solution. The "ring around the rosy method is

I\
y . . = y .. - y . : allj, i=l, ... ,m+l, i*j
lJ lJ rJ

( 1 5)
all j

= Ym+l, O' z . - c.= y +l ., j=l, •.. ,n

J .1 m ,.1

Step 7. Go to step 4, for another iteration. There will be a finite

iterative process to find an optimal solution.

 13

The dual simplex algorithm

Every linear programming problem (primal) has associated

with it a companion problem which is called the dual.

The solution to the dual is simultaneously determined when

solving the primal problem. In many cases, the duality can be of con-

siderable help in solving linear programming problems.

Because the dual simplex algorithm will be used to determine

the vectors to enter and leave basis in integer linear programming,

a short discussion of this algorithm will be presented here.

When the simplex tableau have one or more x . < 0, but with
Bl
z . - c. > 0 for all j, the criteria for the change of basis is;
J J -
1. Vector to be removed from basis: Choose

( 1 6)

column br is removed from the basis and x is therefore

Br
driven to zero.

2. Vector to enter basis: The vector ak to enter the basis is

determined from

Z - C Z - C
k k j j
0 = = max y .< 0 ( 1 7)
yrk J y rj rJ

Note that in the dual simplex algorithm one first determines the vector

to leave the basis and then the vector to enter. This is the reverse of

that done in the simplex method.

 14

INTEGER PROGRAMMING BY CUTTING

PLANES METHODS

Description

An integer programming problem is a linear programming

problem with the additional feature that some or all of the variables

are required to take on integer values only.

The most general problem of this type of problem can be

written as

Ax= b; x > 0; xj an integer, J E J

1
( 18)
max z = ex

if J contains all j, we have an all integer problem. If J is

1 1
empty, ( 18) is a linear programming problem and if J contains part
1
of j, we have an mix-integer problem. In an all integer problem if

x . can take only the value of 0 or 1, we consider it a 0, 1 integer pro-

J
gramming problem.

In theory, every integer linear programming problem can be

solved. However, in practice, due to the difficulties of too much

time involved to obtain the solution (even with today's computer) and

the effect of round-off error and machine storage size, more research

work in this area is still needed.

 15

Formulation methods and solutions

A basic technique for solving the integer programming problems

is (Hadley, 1964).

Step 1. Solve the problem ignoring the integrality restriction.

Step 2. If an optimal solution is not feasible due to the violation of

the integrality restrictions, a new constraint is added to the

original constraint equations, then proceed to step 1. Other-

wise, the optimal solution with the required integer variables

has been obtained.

The first method discussed here is Gomory's Algorithm. The

essential nature of this algorithm is that it provides the proof of finite-

ness.

Gomory's algorithm for the all integer programming problem

Step 1. Solve LP (18) and let the optimal solution so obtained by x = y 0,

8

where

= x +~ y,x. ( 1 9)
8 jER J J

R is the set of j corresponding to the nonbasic variables.

Step 2. If all components of y are integers, we have an optimal

0
solution; otherwise let

xBu = y - Ly .x., ( 20)

uO jER UJ J

where yuO is not integeral.

 16

Now write

( 21 )

where er . is the largest integer< y .,

~ - ~

jER and j = O; and O < f . < 1.

UJ

If more than a single component of y O is

nonintegral, we adopt the equation for which

fuO is largest, where O < fuO < 1.

Step 3. Add constraint (22) to the constraints of LP (16)

f Q - ~ f ,X , < 0 ( 22)
u jER UJ J -

where

fuO > 0, 1 > f . > 0.

UJ -

Step 4. Use the dual simplex algoritrun to obtain an optimal solution

to the augmented problem and let xB= y then return to step 2.

O
Gomory's algoritrun for the mixed-integer programming problem.

Step l. Same as all integer case.

Step 2. Same as all integer case.

Step 3. Use formula (23) and (24) to obtain the cut.

~ (-cl .)x . < -f (23)

jcR UJ J- u0
 17

y . , jER,:+, , for the variables not required to be integer.

UJ

f
1 -f
uO j1

y uj
I '
J.E"R,:, (24)
uO
d .
UJ
f . , jER()J , f .< f , for the variables require to take
UJ 1 UJ- U0

integer value.
f
uO , f
1 - f ( l - f UJ. ) , j E R ( IJ 1 , f . > u O'
uO ~

,,.
,.
where R is the sub~et of R not in R0J ,
1
,,, .,.
_,, ,,,,.
R~ and R are the subsets of R for which

y . > 0 and y . < 0 re spec ti vely.

UJ - UJ
Step 4. Same as all integer cases.

Another method for solving the 0, l all integer programming

problem is suggested by Ferydoon Kianfar. This method uses Gomory's

cutting- plane method to obtain the first cut and then allows the hyper-

plane of that cut to be rotated in any direction until the final hyperplane

passes through as many integer points not in convex set S (defined below)

as possible (Kianfar, 1971 ).

Let the sets of points be

( 25)

and
 18

n I
= ( x , . . . , xn) 1
x E"S: ~ a .x . < L ! (26)
1 I j=l J J- !

w h c r <' a . . a 11 j , and T, a r c- po s i t iv e int e g <'r con st ant s .

J
The problem is to find the largest integer a', r < n, denoted
r
by ~ , sue h that a' > a and A = A , where
r r r r

A = X
r ,
~ l XcS; j~= l , j *r a.x.+a'x
J J r
<L
r-
\
)
( 27)

if we assume that a > a , then the inequality

r r

n
~ a .x . + a/1. x < L (28)
j=l, j;tr J J r r-

is stronger than the original inequality

n
~ a.x.< L ( 29)
j=l J J-

where ~ should be selected such that

r

X = 1) , A cA and A =A for a 1 = a
r 1 r r r r

A
this is equivalent to selecting a as
r

/\ n
a = L -maxX A~ a.x. (3 0)
r E j=l,j;i:r J J

but in fact ~r is equal to a +Sr (L-a

r n r
) by theorem 1

where
 19
r r
S (L -a ) = L- a - F (L-a ) ( 31}
n r r n r

and

n
F: (L -a ) = maxX A l:: a .x . (3 2)
r E 1. 1 ·...1. J J
J= ,J+r

is a special case of the Knapsack Function (Kianfar, 1971 }.

Using the Knapsack Function we have

Theorem 1. let

j n
M = ,, X = (x , .. . ,x )ll:: a.x.< L; x. = 0, 1 , 2, ... , all j ~(3 3}
; 1 n 'j=l J J- J
'
I n
N = ~ X = (x , ...
1 l
, x } .2: .
I a .x.+ (a + k )x < L;
n J= 1, J:#r J J r r r -

X
j
= 0, 1, 2, ... , all j f, (34)

where

r
O~k;:.minn=l, 2 , ... , [ L/a ] (Sn (L-har)/h)
r

Then M = N. (The notation f y] means the largest integer

less than or equal to y.)

Proposition 1.

s1 (k) = oo, k< 0; s 1 (k} = k, O~k<a ;

1

S ( k) = k - a , a < k < L; (3 5)
1 1 1- -

'
S. (k) = min~ S. (k), S. (k-a . )} (j=2, ... , n)
J l J- 1 J-1 J )
 20

Proposition 2.

S. (k) is either zero or is equal to S.(k- 1) + 1.

J J

Where S. (k) = k - k' and k' is the largest integer p < k such that S.(p)=0.
J - J
Now let

then proceed to find D . From (3 5) we obtain

n

D.=D. +jk+a.jkED. ,k+a .<L( ! , (j=2, ... ,n)

J J- 1 ) J J- 1 J-

each D ., j=2, ... ,n is computed from D. by min )N(D. ), L-a . l

J J- 1 ' J- 1 J(
wh ere N(D. ) denotes the number of elements (Kianfar, 1971 ).
J- 1

The Sr (k) of the Knapsack Function is that

n

r r r n
S (k) = k - F (k), wher e F (k) = max :E a.x.
n n n j= 1 , j fr J J

n
subject to :E a.x.= k, x = 0, 1, all j,
j= l , jf r J J j

then

,.
'Jr = a + Sr (L-a ) (3 7)
r n r

and
 21

Sr (k) = k -k', (k=O, 1, ... , L) (38)

n

(
Using (37) and (38) we comp ute a next, we set r = n - 1
n
n -1 n
and compute D but D., j=l, ... , n - 2 and hence
n J

D n-1 = Dn + ~ k+afl I kt:D n J.

n n-2 n , n -2, k+a < L ( 3 9)
n-

From !'roposition 1 it follows that

r n 1/\ n fl }
D
r +l
= D
r-
+ lk+a
1 j
'
r +1
1
I
kt: D
r-1
, k+a < L
r + 1-
( 40)

I .
D~ = D~
J
+ ) kt~. \ kt:D~ l k+~.< L
J- 1 ) J, J- , r J~ , (j=r+2, . .. , n) (41)

The steps used to get a stronger cut is (Kianfar, 1971)

Step l. Apply Gomory's algorithm to obtain original

cut. The cut at this stage is

(42)

where

k
is the vector of the constants of this cut. T has 1 as

its first component and the nonbasic variables of the kth

tableau as the rest of its components.

Also we define
 22
k
F p (43)

-1
where P is an (n+l) x (n+l) matrix whose inverse p

transforms AO into Ak as follows:

k O -1
A = A P ( 44)

The e quivalent cut in 0, 1 variables is then

Ste p 2. Find Dn = D . , j = 1, ... n

,n -1 using (36), and D
j J n

Compute Qr using (3 7 ) and (38) and set r = n - 1, wher e r is

is a variabl e that is to run ov e r the c oordinat e currently

being considered for an increase in its coeffi c ient (Kianfar, 1971 ).

n
St e p 3. If r = 1 go to step 4; also check whether (L -a ) ~ D • If
r r - 1

yes, set~ = a ; otherwise, using (40) and (41) to compute

r r
r f1
D , and use (37) and (38) to find a . In either case, set
n r
r = r -1 and go to step 3.

Step 4. Set D~ = ~O, ~~- Use (41) to compute D:, and use (37) and

(38) to compute a , and we arrived at strongest inequality,

1

the coefficient is denoted by a. for all a.

J J
n
~ a.x . < L ( 45)
j= 1 J J -
 23

NUMERICAL EXAMPLES

In the previous chapter, two methods to solve 0, 1 integer

linear programming problem by cutting plan,es methods have been

discussed. It will be helpful to illustrate the two methods by consider-

ing some simple numerical examples. As the first example, assume

the following 0, 1 integer linear programming problem:

E x ample 1.

subject to

2x + 4x - 4x + 2x + 4x ~ 7
1 2 3 4 5

xl + 3x + 2x 3 + X - x5 ~ 4
2 4
(46)
2x - 2x - x4 + x5 ~ 3
2 3

x . =0orl, J=l, ... ,5

J

Table 1. The initial primal feasible tableau.

X
,B 1 xl x2 ~3 x4 ~5
z 0 -7 8 9 -4 -5
x6 7 2 4 -4 2 4
x7 4 1 3 2 1 -1
XS 3 0 2 -2 -1 1
 24

Table 1. Continued

x9 1 1 0 0 0 0

xlO 0 1 0 0 0
xl I 0 0 l 0 0
xl2 l 0 0 0 l 0
xl3 l 0 0 0 0 l

Table 2. The optimal linear programming tableau with no cut.

X x2 x3 x6 x9
XB xl2

z 14.75 13.00 4.00 1. 25 4.50 1. 50

x4 1. 00 0.00 0.00 0.00 0,00 1. 00

x7 2. 75 4.00 1. 00 o. 25 -1. 50 -1. 50

x8 3. 2 5 l. 00 -1. 00 o. 25 o. 50 1. 50
X l. 00 0,00 0.00 0.00 l. 00 0.00
l
xlO 1. 00 1. 00 o.oo 0.00 0.00 o. 00
xll l. 00 o. 00 1. 00 0.00 0.00 0.00

xl3 0.25 -1. 00 l. 00 -0. 25 0.50 0. 50

XS o. 75 1. 00 -1. 00 o. 25 -0. 50 -o. 50

First , to solve this problem by the regular simplex method

without considering its integer requirement.

The solution is given in Table 2. Notice the optimal solution

does not meet the integrality restriction.

To obtain the first cut, using the (x = x ) row in Table 1.

B2 7
Note that y"21J= 2. 75, Yzz = 4. 00, y
23
= 1. 00, y
26
= 0. 25, y
29
= -1. 50,
 25

y
212
= -1. 50. We see that £
20
= 0. 75, £
22
= 0, £
23
= 0, £
26
= 0. 25,

£29 = o. 5, £2.12= o. 5
Thus by (22), the cut is

0. 75 - 0. 25x - 0. 5x - 0. 5x ~ 0 ( 4 7)
6 9 12

Annex (47) to the original problem and apply the dual simplex

algorithm to solve it.

One obtains an optimal solution in two iterations. For this

solution given in Table 3, all x are integers. Thus, one has obtained

the optimal solution to ( 46 ). It is x = x = 1, x =x =x = 0, z = 1 2.

1 5 2 3 4

Table 3. The optimal tableau after the first Gomory 1 s cut

xB 1 xz x3 x4 x9 xsl

z 1 2. 00 13.00 4.00 1. 00 2.00 5.00

x6 1. 00 0,00 0.00 -2. 00 2.00 -4. 00

x7 4.00 4. 00 1. 00 2. 00 -2. 00 1. 00
X 1. 00 0.00 0.00 0.00 1. 00 0,00
1
xlO 1. 00 1. 00 0.00 0.00 0.00 0 .. 00
0.00 -1. 00 1. 00 -1. 00 1. 00 -1. 00
xl3
XS 1. 00 1. 00 -1. 00 1. 00 -1. 00 1. 00

xl2 1. 00 0.00 0.00 1. 00 0.00 0.00

 26

Next one considers to use Kianfer's stronger cut, which modify

the Gomory's cut. From (47) one has

0. 75 - 0. 25x - 0. 5x - 0. 5x f 0
6 9 12

or

FT= (0. 75, 0, 0, - O. 25, -0. 5, -0. 5), [ 1, x , x , x , x , x ] '.5._0

2 3 6 9 12

to convert this cut to a cut that contains only 0, 1 variables, cme first

finds

1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
P= 7 -2 -4 4 -2 -4
1 -1 0 0 0 0
1 0 0 -1
...
0 0
J

0
Then F = FP = (-2, 1, l, -1, 1, 1)

The equivalent cut with 0, 1 variables

or

(48)
 27

We make this inequality stronger by the method of replacing

inequalities;

To c ompute ~S we comput e D~ (See Table 4),

From (37), (38), it follows that s!(L -a ) = S~(l) = 1 - 1 = 0,

5
and { = l+O = 1. To find
5 4
1
we first check to see whether (L-a
4
)
5 5 A
E D , and L-a = 2-1 :: 1 E D , thus a = a = 1. To find a3' we
3 4 3 4 4
5 3 .
c h ec k L - a = 2-(-1)
3
= 3 D ; hence we compute
2
1
D using (40),
5
3 3
(41) (S e e Table 5), from (37), (38) we have S (L-a ) = s (3) = 3-2= 1,
5 3 5
f· r\
a = -1 +1 = 0. Next we compute a .
2
First we check L -a = 2-1 = 1
3 2
5 A 1
c D , h e nce a = a = 1. One now computa:.D as shown in Table 6,
2 2 2 5
1 1
us i ng (38), we have S (L-a ) = S (1) = l - .1 = 0, Hence, from (37)
5 1 5

Table 4. Computation . o£ D5
5

5 5 5 D5=D5
Dl D2 D3 4 5
0 0 0 0
1 1 1 1
2 2 2
 28
3
T a bl e 5 . Computation of D
5

3 3 5 3
D :5= D l o2 D5

0 0

1 l
2 2

T a bl e 6 . Computation . o f D_l
- :,

1 l 1 Dl
D2 D3 D4 5
0 0 0 0
1 l 1 1
2 2

Th e modified cut now is F o= (- 2, 1, 1,0, 1, 1) or

( 49)

-1
To compute F we need p , which can be obtained from Table 2.
(
i 1. 00 0,00 0.00 0.00 0.00 0. 00 I
!
I 1. 00 0.00 0.00 0.00
l -1. 00 0.00
-1 0.00 1. 00 0.00 0.00 0.00 o. 00
p =
0.00 0.00 1. 00 0.00 0.00 0, 00 I

1. 00 0.00 0.00 0.00 0.00 -1. 00

j
I

l 0.75 -1. 00 1. 00 -0. 25 0.50 o. 50 ..i

 29

The vector F is
-1
F = F P = (0. 75, 0, 1, -0. 25, -0. 50, -0. 50)
0

and the stronger inequality in terms of nonbasic variables of the

tableau of Table 2 is

0. 75 + x3 - 0. 25x
6
- 0. 5x - O. 5x ~ 0
9 11

This row is annexed to Table 2 and the resulting tableau is

made optimal by the dual simplex method.

Table 7 shows the optimal tableau. For this solution , the integer

requirements are satisfied. Thus one has obtained the optimal solu-

tion to (46). It is

xl = X = 1, X = X = X = 0, z = 1 2.
5 2 3 4

Table 7. The optimal tableau after the first Kianfar's cut

XB x2 x3 x4 x9 X
sl
z 12. 00 13 . 00 9.00 1. 00 2.00 5.00

x6 1. 00 0.00 -4. 00 -2.. 00 2.00 -4. 00

x7 4.00 4.00 2.00 2.00 -2. 00 1. 00

x8 2.00 1. 00 -2. 00 -2. 00 1. 00 -1. 00

X 1. 00 0.00 0.00 0.00 1. 00 0.00
1
xlO 1. 00 1. 00 0.00 0.00 0.00 0.00

xll 1. 00 0.00 1. 00 0.00 0.00 0.00

xl3 0.00 -1. 00 0.00 -1. 00 1. 00 -1. 00

x5 1. 00 1. 00 0.00 1. 00 -1. 00 1. 00

xl2 1. 00 0.00 0.00 1. 00 0.00 o. 00

 30

In addition to the above example, three more examples are

provided below.

Example 2.

subject to

x . = 0 or 1, j = 1, ... ,5
J

Using Gomory's algorithm, one ha.s the following results .

Initial solution Z = 5. 77
1st cut Z = 5. 42
2nd cut Z = 5. 00

After two cuts one obtaiQ;;the integer solution with

Using Kianfar's algorithm , one has the following results.

Initial solution Z = 5. 77
1st cut Z = 5. 00

The solution is the same as in the Gomory's algorithm,

except this method only needs one cut to obtain an integer solution,

The third problem is from Kianfar's · example.

 31

Example 3,

Max. 1 5x
1
- 2x + l 6x + 6x
2 3 4

subject to

3 6x 46x + 18x 34x < 28

1 2 3 4 -
75x
1
20x
2
+ 30x
3
+ 27x
4
<
- 93

78x
1
+ 9lx
2
+ 69x
3
16x
4
< 122

x. = 0 or 1, j=l, ... ,4
J

Using Gomory's algorithm the results are

Initial solution z = 20.69

1st cut z = 20.68
2nd cut z = 20.66
3rd cut z = 20.64
4th cut z = 20.62
5th cut z = 20.61
6th cut z = 20. 59
7th cut z = 20. 50
8th cut z = 20.38
9th cut z = 20.37
10th cut z = 20.35
llthcut z = 20.34
12th cut z = 20. 28
13th cut z = 20. 26
14th cut z = 20.22
15th cut z = 20. 21
16th cut z = 20. 21
17th cut z = 20. 20
18th cut z = 20.19
 32

l 9th cut z = 20.15

20th cut z = 19.95
21st cut z = 19. 87
22nd cut z = 19. 85
23rd cut z = 19.86
24th cut z = 19.85
25th cut z = 16. 00

After 25 cuts the solution is x

1
= x 2 = x 4 = 0. x
3
= 1, z = 16
and i s all integers. The above tells us that in this problem Gomory's

algorithm needs 25 cuts to obtain the solution.

Now, by using Kianfar's algorithm the results are:

Initial solution Z = 20. 69

1st cut Z = 19.48
2nd cut Z = 1 6. 00

Here the Kianfar's algorithm uses only two cuts to reach the

solution.

The last and the largest example is presented below:

Example 4.

10x 1 -7x
2
+ x -12x
3 4
+ 2x
5
+ 8x -3x - x + Sx + 3x = Max
6 7 8 9 10

-3x -12x +8x + x

1 2 3 4

x +10x + x - x + 7x + x
2 3 5 6 7 8
< 13

Sx
1
+ 3x - x + 2x 6
2 3 8

4x - 2x
3 4 < 8
 33

= < 12

8x + 5x - 2x - 7x + x .. 5x < 16
1 2 3 4 5 7

x . = 0 or 1, (j = 1, ... , 10)
J

In this example, Gomory's algorithm used 88 cuts and still

did not obtain the solution (Z = 24. 12) while Kianfar's algorithm

used only 13 cuts to get the solution.

The solution is

=X9 = 0

z = 23.
 34

COMPARISON OF THE EFFECIENCY, USAGE,

AND RELATIONSHIP OF DIFFERENT

CUTTING PLANES METHODS

All the four numerical examples cited in the previous chapter

have been tested by computer codes generated by the author using

Fortrain IV language on B 6700.

From the four ex ampl es discussed, it can be seen that the

number of cuts needed to obtain the solution by Gomory's method are

greater than or e qual to Kianfar 's method in 0, 1 integer programming

problem. Kianfar I s method may be viewed as the modification of

Gomory's method in actual practice. The only difference between

these two methods is that after the cut the Gomory's method has been

obtained, Kianfar 's method changes that cut so as to make it stronger.

It is believed that the larger the problem the more efficiency

the Kianfer 's method can be seen. The results from Kianfer 1 s method

are very encouraging but of course this experience is insufficient for

a firm judgement of the efficiency of the algorithm. However, Gomory 1 s

algorithm imposes a heavy burden on the storage system and processing

time on a computer, if the problem is large. Kianfar 1 s method tends

to have the ability of lessening this burden.

Due to the fact that the computer programs and the numerical
 35

examples given are problem specific, the comment made here can

only be viewed as a suggestion or reference. In addition, the

efficiency of each method is partly dependent on the nature and

characters of the problem being solved.

Summary

In summary the advantage and disadvantage of the two methods

will be discussed.

The Gomory's method: This method has the ability to solve

all kinds of all integer and mix-integer programming problems. It is

e asier to learn and perform, and more suitable for small problems.

Less arithmetic operations are needed to form a cut. However, if

the problem turns large, the nurn ber of cuts needed to solve the problem

increases rapidly.

The Kianfar 1 s method: This method can be used to solve 0, 1

integer programming problem only. It is hard to learn and not

efficiency on small problems . Usually more arithmetic operations

are needed to form a cut, however if the problem turns large the

number of cuts needed to solve the problem increases gently.

Another point that is worth mentioning is that the Kianfar 's

approach can also be applied to the original constraints. In order to

maintain the finiteness property of the Kianfar's algorithm, Gomory's

cut should be added to it every 50 or 100 iterations.

 36

Table 89-· Gomory's algorithm

Pro.)lem No. of No. of Virtual Processor No. of No. of

No. variables constrnts memorx: time cuts Ite.
1 5 4 1. 1 7 o. 12 l 6
2 5 4 1. 34 0. 14 2 5
3 4 4 1. 64 0. 22 25 48
4 ,:,,~ 10 7 41. 22 3. 11 88 299

Table Sb. Kianfar's algorithm

Problem No 0f No. of Virtual Processor No. of No. of

No. variables constrnts memory time cuts Ite.

1 5 4 2.69 o. 25 1 6
2 5 4 2. 79 0. 24 1 4

3 4 4 2. 58 o. 25 2 8
4 10 7 3. 59 0.42 13 50

,:, Th(• unit of virtual memory 1s kiloword-minutes.

,:, The unit of processor time is minutes.

,:, Both compilation time and execution time are all included in memory
and time calculation.
,:n:,Problem No. 4 by Gomory's algoritlun after 88 cuts still does not
obtain the answer.
 37

REFERENCES

Balas, Egon. 1965. "An Additive Algorithm for Solving Linear

Programs with Zero-one Variables, 11
Operations Research
13: 517-546.

Balinski, M. L. 1965. "Integer Programming; Methods, Uses,

Computation," Management Science, 12(3):253-313.

Geo ffri o n, A. M . and R . E. Marsten. 1972. "Integer Programming

Algorithms; A Framework and State-of-the-Art Survey,"
Manag e ment Science, 18(8)461-491.

Hadl e y , George. 19 6 2 . "Lin e ar Programming." Massachusetts,

Addison-Wesley, 520 p.

1964. "Nonlinear and Dynamic Programming. 11

Massa c husetts, Addison-Wesley, 484 p.

Kianfar, Ferydoon. 1971. "Stronger Inequalities for Integer Pra-

g ramming Using Knapsack Functions 11
0 perations Research,
19:1374-1392.

Kuo, Shan Sun. 1965. "Numberical Analysis" Massachusetts,

Addison-W e sley, 341 p.

Trauth, C. A. and R. E. Woolsey. 1969. "Integer Linear Programming :

A Study in Computational Efficiency," Management Science
15(9) :481-493.
 38

APPENDIX
 39

Computer Programs

C THIS PROGRAM USE GO MO RY'S ALGORITHM TO SOLVE Al L

C INTE G ER LINEAR
C PROC ~RAMMING PROBLEM. WHERE II = TOTAL NUMBER OF
C RO WS.
r,J LD( I)=VARIA BLES NOT IN THE BASIS.
C B( I) = THE PIVOT ROW.
C L( I)=VARIA BLES IN THE BASIS.
C J J =TOTAL NUMBER OF C OLUMNS. S(I, J)=TABLEA U FOR
C SIMPLEX METHOD.
DIMENSION S(80, 180), L(80), DUJ(80), LD(80) , B(80)
R E A D ( 5 , 4 ) II , J J
4 FORMAT( 1814)
III.= II+l
C READ IN THE MA TRIX
DO l O I= l , III
lO R EA D ( 5 , 1 2 ) ( S (I, J ) , J = 1 , J J )
12 FORMAT ( 7 F 1 0. 4)
C READ IN THE SUBSCRIPT FOR THE SLACK VARIABLE
READ(S, 4) (L(I), 1=2, II)
IIII=II-1
REGA=0.
21 KKK=0
REGB=0.
C CALCULATE NEW LAST ROW
22 DO 23 1=2, II
IF(L(I), NE. 0) GO TO 23
DO 27 J=l, J J
IF(S(I, J). EQ. 0.) GO TO 27
S(III, J)=S(III, J)-S(I, J)
27 CONTINUE
23 CONTINUE
C SEARCHING FOR THE MOST NEGATIVE ENTRY
K=III
44 JJJ=JJ-1
IF(REGA. EQ. l. ) GO TO 500
B(K)=0.
L(K) = 0
DO 4 2 J = 1 , J J J
IF(S(K, J ). GE. 0.) GO TO 42
IF(B(K), LE. S(K, J) ) GO TO 42
 40

B(K)=S(K, J)
L(K)=J
42 CONTINUE
C TEST FOR L(K).
REGB=l.
45 IF(L(K)) 46, 62, 46
C FIND OUT THE PIVOT COLUMN
4 6 KJ=L(K)
DO 120 I=2, II
IF (S (I, K J)) 1 2 0, 1 2 0, 1 21
1 20 CONTINUE
WRITE(6, 130)

130 FORMAT(lX, 'UNBOUNDED')

GO TO 1
C COMPUTTING THE RA TIO
121 JK=0
DO 50 I=Z, II
IF(S(I , KJ), LE. 0.) GO TO 50
X=S(I, J J)/S(I, KJ)
IF(JK. EQ. 0) GO TO 53
IF( X. GE. XMIN) GO TO 50
53 XMIN=X
JK== I
5 0 CONTINUE
5 6 X = S ( J K, KT)
L(JK)=KJ
C CALCULATING THE NEW ROWS A BOVE THE PIVOT ROW
DO 57 .I=l, III
5 7 B(I)=S(I, KJ)
IJ=JK-I
DO 59 I= 1, IJ
DO 59 J=l, J J
IF (S ( J K, J)) 5 8 , 5 9, 5 8
58 IF (B(I)) 580, 59,580
580 S(I, J) = S(I, J) - B(I) ':'(S(JK, J)/X)
59 CONTINUE
C CALCULATING THE NEW ROWS BELOW THE PIVOT ROW
IJ=JK+l
DO 61 I=IJ, III
DO 61 J=l, J J
IF (S(JK, J)) 60, 61, 60
60 IF(B(I))600, 61,600
600 S(I, J)=S(I, J)-B(I)*(S(JK, J)/X)
61 CONTINUE
C NORMALIZATION
 41

DO205J=l,JJ
205 S(JK, J)=S(JK, J)/X
KKK=KKK+l
WRITE(6, l 05) KKK, S(K, JJ), L(JK)
105 FORMAT (1X,I4,6X,Fl5.2,10X.I4)
GO TO 44
62 IF (K-1) 70, 70, 63
63 IJ=JJ-1
C TEST TO SEE WHETHER ALL THE ELEMENTS ON THE
C LAST ROW ARE CLOSE TO ZERO.
DO 65 J=l, IJ
IF(S(K, J)-. 0001) 65, 65, 66
65 CONTINUE
WRITE (6,103)
103 FORMAT (lX, 'FE ASIBLE')
WRITE(6, l 01)
101 FORMAT(lX, 'ITERATION OBJ. FUNCTION NEW BASIC VAR.'
C TEST TO SEE WHETHER ALL THE ELEMENTS ON THE LAST
C ROW
DO140J =l,JJ
140 S(III), J)=0.
K=l
KKK=0
IF(REGB. EQ. 0) GO TO 31
GO TO 44
31 REGB=l.
GO TO 500
66 WRITE(6, 6)
6 FORMA T(l X, 'INFEASIBLE')
GO TO l
70 WRITE(6, 8) S(l, JJ)
8 FORMAT(/// /lX, '0 BJ. FUNCTION,', F20. 8/, lX)
C FIND THE LARGEST S(I, JJ)
199 FUO=0.
DO 202 I= 2, II
REMAN=S(I, JJ)-AINT(S(I, JJ))
IF(REMAN. GE. 0. 9999995) GO TO 201
IF(REMAN. LE. 0. 0000005) S(I, JJ)=AINT(S(I, JJ))
IF(REDMAN. LE. FUO) GO TO 202
FUO=REMAN
LL=I
GO TO 202
201 S(I, JJ)=S(I, JJ)+0. 0000005
S(I, JJ)+AINT(S(I, JJ))
202 CONTINUE
IF(FUO. LE. 0. 5E-5) GO TO l
 42

C
3 00 INTS=JJ-II
DO 216 I=l, INTS
216 DUJ(I)=O.
JC=l
DO 21 8 I= 1 , J J
DO 220 J= 2, II
220 IF ( I. EQ . L ( J)) GO TO 21 8
LD(JC )= I
JC=JC+l
218 CONTINUE
JC=JC-1
DO 21 2 J = l , JC
NOTB=LD(J)
IF(S(LL,NOTB).LT. 0. )GO TO 213
FUJ =S(LL, NO TB)-AINT(S(LL, NO TB))
GO TO 212
213 ABSY=ABS(S(LL, NOTB))
FUJ:=A BSY-AINT(A BSY)
FUJ=l. -FUJ
21 2 DUJ( .T)=FUJ
C INSERT THE NEW CUT INTO MATRIX
JJJ=JJ+l
DO 226 I=l, II
226 S(I,JJJ)=S(I,JJ)
DO 228 I=l, III
228 S(I, J J)=O.
S(Ill,JJ)=l.
S(III, JJJ)=-FUO
DO 230 I=l, INTS
NEWCUT=LD(I)
230 S(III, NEWCUT)=-;JUJ(I)
L(III)=J J
JJJ=JJ
JJ=JJ+l
II=II+l
III=II+l
IF(III. EQ. 50) GO TO l
DO 232 I=l, JJ
232 S(III,I)=O.
REGA=!.
GO TO 21
C MIN XBI
500 JK=O
IF(REGB. EQ. 0.) GO TO 62
XMIN=O.
 43

DO 3 0 2 1 = 2, II
IF(S(I, JJ). GE. XMIN) GO TO 302
XMIN=S(I, J J)
JK=I
3 02 CONTINUE
IF(XMIN. GE. 0.) GO TO 70
IF(XMIN. LE. 0SE-4) GO TO 70
KJ=0
XMAX=-0. 5El 2
DD 304 J=l, JJJ
IF(S(JK, J). GE.0.) GO TO 304
X=S(l , J)/S(JK, J)
IF(X. LE. XMAX) GO TO 304
306 XMAX=X
KJ=J
304 CONTINUE
IF(KJ. EQ. 0) GO TO 1
GO TO 56
1 WRITE(6, 100)
l 00 FORMAT(////' THE FINAL MATRIX')
DO 78 I= l,III
WRITE(6, 1 50) I
l 50 FOR MA T (/ / 3 5 X, 1 RO W ', I 2 /, 1 X)
78 WRITE(6, 12) (S(I, J), J=l, J J)
WRITE(6, 7)
7 FORMAT (lX, 'VARIABLE VALUE')
DO711=2,II
71 WRITE (6, 5) L(I), S(I, J J)
5 FORMAT (I4, F20. 8)
STOP
END
 44

C THIS PROGRAM USE THE ALGORITHM SUGGESTED BY

C DR. FERYDOOM KIANFAR
C FOR SOLVING 0, l ALL INTEGER PROGRAMMING PRO-
C BLE MS USING KNAPSACK
C FUNCTIONS. WHERE !!=TOTAL NUMBER OF ROWS.
C JJ=TOTAL NUMBER OF COLUMNS. S(I, J)=TABLEAU FOR
C SIMPLEX METHOD.
C L(I)=VARIABLES IN THE BASIS.
C B(I)=THE PIVOT ROW.
C LD(I)=V ARIA BLES NOT IN THE BASIS.
C FT(I)=TH.E F VECTOR.
C SS(I, J)=THE INITIAL PRIMAL FEASIBLE TABLEAU.
C P(I, !)=THE P MA TRIX.
C FP(I)=THE VECTOR COME FROM P MATRIX MULTIPLY
C BY FT VECTOR.
C IDA(I)=TABLEAU CONTAINING THE VALUE OF D.
C IDB(I)=TABLEAU CONTAINING THE VALUE OF D NOT IN IDA.
C IARH(I)=NEW VALUE OF A.
C IDRR=INDEX VALUE FOR TABREAU IDA(!) AND IDB(!).
C IFP(I)=INTEGER VALUE OF FP(I).
C NARI(I)=THE LEGITIMATE VARIABLES.
DIMENSION S(46, 96), B(46), L(46), FT(30), LD(30)
DIMENSION IDB(300~ IDRR(30), FP(30), NARI(l 5)
DIMENSION SS(46, 96), P(30, 30), IFP (30), IDA(300), IARH(30)
R EA D ( 5 , 4) II, J J
4 FORMAT ( 18 !4)
III=II+l
C READ IN THE MA TRIX
DO l 0 I= l , III
10 READ(5, 12) (S(I, J), J= l, J J)
12 FORMA T(7Fl 0. 4)
DO 3 l =l, III
DO 3 J=l, J J
3 SS(I, J)=S(I, J)
C READ IN THE SUBSCRIPT FOR THE SLACK VARIABLE
READ(5, 4) (L(I), !=2, II)
C READ IN VARIABLES THAT NEED TO BE INTEGER
READ(5, 4) NN, (NARI(!), I=l, NN)
IIII=II-1
REGA=0.
21 KKK=0
22 DO 23 !=2,n
IF(L(I). NE. 0) GO TO 23
DO 27 J=l, J J
IF(S(I, J), EQ. 0. ) GO TO 27
S(III, J)=S(III, J)-S(I, J)
 45

27 CONTINUE
23 CONTINUE
C SEARCHING FOR THE MOST NEGATIVE ENTRY
K=III
44 JJJ=JJ-1
IF(REGA. EQ. 1. ) GO TO 500
B(K)=0.
L(K)=0.
DO 42 J=l, JJJ
IF(S(K, J). GE. 0.) GO TO 42
IF(B(K). LE. S(K, J)) GO TO 42
B(K)=S(K, J)
L(K)=J
42 CONTINUE
C TEST FOR L(K).
45 IF(L(K)) 46,62,46
C FIND OUT THE PIVOT COLUMN
46 KJ=L(K)
DO 120 I=2, II
IF(S(I, KJ)) 120, 120, 121
120 CONTINUE
WRITE(6, 130)

130 FORMAT(lX, 'UNBOUNDED')

GO TO 1
C COMPUTING THE RATIO
1 21 JK=0
DO 50 I=2, II
IF(S(I, KJ). LE. 0. ) GO TO 50
X=S(I, J J)/S(I, KJ)
IF(JK. EQ. 0) GO TO 53
IF(X. GE. XMIN) GO TO 50
53 XMIN=X
JK=I
50 CONTINUE
56 X=S(JK, KJ)
WRITE(6, 149) X
149 FORMAT(/,'X= ',Fl4.6)
L(JK)=KJ
C CALCULATING THE NEW ROWS ABOVE THE PIVOT ROW
DO 5 7 I= l , III
57 B(I)=S(I, KJ)
IJ=JK-1
DO 59 I=l,IJ
D059J=l,JJ
IF(S(JK, J)) 58, 59, 58
 46

58 IF (B(I))580, 59,580
580 S(I, J)=S( I, J) - B(I)'~(S(JK, J)/X)
59 CONTINUE
C CALCULATING THE NEW ROWS BELOW THE PNOT ROW
IJ=JK+l
DO 61 I=IJ, III
DO 61 J=l, J J
IF (S ( J K, J)) 6 0, 61 , 6 0
60 IF (B(I))600, 61,600
6 00 S(I, J) =S(I, J}--B(I);'(S(JK, J)/,C)
61 CONTINUE
C NORMALIZATION
DO 2 0 5 J = 1 , J J
2Q5 S(JK, J)=S(JK, J)/X
KKK=KKK+l
WRITE(6, 105) KKK, S(l, JJ), L(JK)
105 FORMAT (lX, I4, 6X, Fl 5. 2, 1 OX, I4)
GO TO 44
62 IF(K-1) 70, 70, 63
63 IJ=JJ-1
C TEST TO SEE WHETHER ALL THE ELEMENTS ON THE LAST ROW
C ARE CLOSE TO ZERO.
DO 65 J=l, IJ
IF(S(K, J)-. 0001) 65, 65, 66
65 CONTINUE
WRITE (6,103)
103 FORMAT (lX, 'FEASIBLE')
WRITE(6, 101)
101 FORMAT(lX, 'ITERATION OBJ.FUNCTION NF-W BASIC VAR.
C TEST TO SEE WHETHER ALL THE ELEMENTS ON THE LAST ROW
DO 1 4 0 J = 1 , J J
140 S(III, J)=0.
K=l
KKK=0
GO TO 44
66 WRITE(6, 6)
6 FORMAT(lX, 'INFEASIBLE')
GO TO 1
70 WRITE(6, 8) S(l, J J)
1
8 FORMAT(/ / / / 1 X, '0BJ. FUNCTION. , F 2 0. 8 /, 1 X)

WRITE(6, 7)
7 FORMAT (IX, 'VARIABLE VALUE')
DO 71 I=2, II
71 WRITE (6, 5) L(I),S(I, JJ)
5 FORMAT(I4 ,F2 0.8)
C FIND THE LARGEST S(I, JJ)
 47

FUO=0.
DO 202 I=2, II
REMAN=S(I, J J)-AIN T(S( I, J J))
IF (REMAN. GE. 0. 9999995) GO TO 201
IF (REMAN. LE. 0. 0000005) S(I, JJ)=AINT(S(I, JJ))
IF RE MAN. LE. FUO) GO TO 202
FUO=REMAN
LL=I
GO TO 202
201 S(I, J J)=S(I, J J)+0, 0000005
S(I, J J)=AINT(S(I, J J))
202 CONTINUE
C TEST IF OPTIMAL SOLUTION HAS BEEN OBTAINED
IF(FUO. LE. 0. 5E-5) GO TO 1
WRITE(6, 954) J J, FUO, LL
9 54 FORMAT(1X, 1 JJ= 1 ,I6,' FUO= 1 ,Fl0.4, 1
LL= 1 ,I6)
300 INTS=JJ-II
DO 216 I=l, INTS
216 FT(I)=0.
JC=l
DO 21 8 I= 1 , J J
DO 220 J= 2, II
220 IF(I.EQ.L(J))GO TO 218
LD(JC)=I
JC=JC+l
218 CONTINUE
JC=JC-1
INTSS=INTS+l
P(l, 1 )=l.
DO 502 J=2. INTSS
502 P(l, J)=0.
DO 504 J=l, JC
:,,roT B=LD( J)
IF(NOTB. LE. NN) GO TO 503
DO 506 1=2,II
506 IF(SS(I, NOTB). EQ. 1.) GO TO 508
WRITE(6, 510)
510 FORMAT(/' ERROR INS MA TRIX')
508 JT=J+l
DO 51 2 K = 1 , INT S
KT=K+l
51 2 P( JT, KT)=-SS(I, K)
P(JT, 1 )=SS(I, J J)
GO TO 504
503 NOT J=J+l
NOTB=NOT B+l
 48

DO 505 MI=l, INTSS

505 P(NOTJ, MI)=0.
P(NOTJ, NOTB)=l.
504 CONTINUE
C COMPUTE FXP
FT(l )=FUO
DO 51 7 I=2, INTSS
IT=I-1
IF(S( LL, LD(IT) ). LT. 0. ) GO TO 513
REMAN=S(LL, LD(IT))-AINT(S(LL, LD(IT)))
GO TO 516
513 ABSY=ABS(S(LL, LD(IT)))
REMA .N=ABSY-AINT(A BSY)
REMAN=l. -REMAN
516 FT(I)=-REMAN
51 7 IF(A BS(FT(I)). GE. 0. 999995) FT(I)=0.
WRITE(6, 916)
916 FORMAT (' FT VECTOR')
WRITE(6, 12) (FT(I), I=l, INTSS)
C FXP
DO 51 8 I= 1 , IN TSS
518 FP(I)=0.
DO 521 I=l, INTSS
DO 520 J=l, INTSS
5 20 FP(I)=P(J, I) ':' FT(J)+ FP(I)
IF(FP(I). GE. 0.) IFP(I)=FP(I)+0. 5
5 21 IF(FP(I). LT. 0. ) IFP(I)=FP(I)-0. 5
LB=-(FP(l )-0. 5)
WRITE(6, 950) LB, FB(l)
1
95 0 FORMAT(lX, LB=',I4, 1 FP(l)= 1 ,Fl0.4)

WRITE( 6,938)
938 FORMAT(' FP(I)')
WRITE(6, 12) (FP(I), I=l, INTSS)
WRITE(6, 12) (IFP(I), I=l, INTSS)
INR=INTSS
C FIND Dl AND DJ, J=2, ... , N
IDA(l)=0
IDA(2)=IFP(2)
KC=l
IDRR(l )=2
K=3
MM=2
J=3
53 8 I= 1
536 IF(I. GT. MM) GO TO 532
NSTRAG=IDA ( I)+ IFP(J)
 49

IF(NSTRAG. GT. LB) GO TO 534

IDP=K-1
DO 535 ICC=l, IDP
535 IF(NSTRAG. EQ. IDA(ICC))GO TO 534
IDA(K)=NSTRAG
K=K+l
534 I=I+l
GO TO 536
532 MM=K-1
J=J+l
KC=KC+l
IDRR(KC)=MM
IF(J. NE. INTSS) GO TO 538
C FIND AR HEAD
KLA=LB-IFP(IN'TSS)
KSNR=KLA-KPLAN(KLA, IDA, K)
WRITE(6, 918) KLA, LB, IFP(INTSS), KSNR
1 1
918 FORMAT(2X, KLA= , I4,' LB=', I4, 1 IFP(INTSS)=', I4,
l 'KSNR=', I4)
IN=INR-1
IAR H(IN)=IFP(INR )+KSNR
IR=IN-1
638 IF(IR. EQ, 1) GO TO 603
MINUSR= IR-1
IPLUSR=IR+l
IEADAK=LB-IFP(IPL USR)
DO 540 I=l, IDRR(MINUSR)
540 IF(IEADAK. EQ. IDA(I)) GO TO 641
C COMPUTE DRN AND FIND AR HEARD
DO 542 I=l, IDRR(MINUSR)
542 IDB(I)=IDA( I)
I= 1
K=IDRR(MINUSR)+l
MM=IDRR(MINUSR)
INR=IN-IR
CALL KNAP(INR, MM, K, IDB, IARH, LB, I, IPLUSR, IFP, &770, IR)
GO TO 638
641 IAR H(IR )=IFP(IPL USR)
IR=IR-1
GO TO 638
C STEP 3
603 IDB(l )=0
MM=2
K=3
I= 1
IPLUSR=3
 50

IDB( 2)=IAR H( 2)
INR=IN-2
CALL KNAP{INR, MM, K, IDB, IARH, LB, I, IPLUSR, IFP, &770, IR)
GO TO 638
770 DO 64 2 I=2, INTSS
IOU=I-1
642 FT(I)=IARH(LOU)
FT(l )= FP(l)
C GET P-1
DO 704 I=l, NN
IK=I+ 1
DO 7 0 6 J = 2 , II
706 IF{L(J). EQ. NARI(!)) GO TO 708
DO 714 K=l, INTSS
714 P(IK, K)=0.
P ( IK, IK) = 1 .
GO TO 704
708 DO 712 K=l, INTS
KT=K+l
71 2 P(I K, KT)=-S(J, LD(K))
P(IK, 1 )=S(J, J J)
704 CONTINUE
C MA TRIX FOP-1
DO 718 I=l, INTSS
71 8 FP(I)=0.
DO 720 I=l, INTSS
DO 720 J = 1, INTSS
720 FP(I)=P(J, l) ':' FT(J)+FP(I)
WRITE (6,930)
930 FORMAT(' FOP-1 ')
WRITE(6, 12) (FP(I), I=l, INTSS)
C INSERT THE NEW CUT INTO MATRIX
JJJ=JJ+l
DO 226 I=l, II
SS(I, JJJ)=SS(I, JJ)
226 S(I, JJJ)=S(I, JJ)
DO 228 I=l, III
SS(I, J J)= 0.
228 S(I, JJ)=0.
S(III, JJ)=l.
SS(III, JJ)=l.
S(III, JJJ)=:-FP(l)
SS(III, J J J)=-:F T(l)
DO 23 0 I=l , INTSS
WRITE (6, 967) INTS, I, LD(I), FT(I)
96 7 FORMAT( 14, 14, I4, Fl 6. 8)
 51

NE WC UT=LD{I)
IQ=I+l
SS(III, I)=FT{IQ)
230 S{III, NEWCUT)=FP(IQ)
L(III)=J J
JJJ=JJ
JJ=JJ+l
II=II+l
III=II+l
IF(III . GE. 25) GO TO 1
DO 23 2 I= l , J J
23 2 S{III, I)=0.
REGA=l.
GO TO 21
C MIN XBI
500 JK=0
XMIN=0.
DO 302 Iz:2, II
IF(S(I, JJ). GE. XMIN) GO TO 302
XMIN=S(I, JJ)
JK=I
302 CONTINUE
C TEST XBR<0
WRITE(6, 997) XMIN
997 FORMAT{' XMIN= ',Flo. 6)
IF(XM!N. GE. 0.) GO TO 70
KJ=0
XMAX=-0. 5El 2
DO 304 J=l, JJJ
IF(S(JK, J). GE. 0.) GO TO 304
X=S(l, J)/S(JK, J)
IF(X, LE. XMAX) GO TO 304
306 XMAX=X
KJ=J
3 04 CONTINUE
IF(KJ. EQ. 0) GO TO 66
GO TO 56
1 WRITE{6, 100)
100 FORMAT(//// 1 THE FINAL MA TRIX')
STOP
END
 52

SU BROUTJJ\rE KNAP( INR, MM, K, IDB, IARH, LB, I, IPLUSR, IFP, *, IR)
DIMENSION IARH(30), IDB(300), IFP(30)
DO 631 J=l, INR
636 IF(I.GT.MM) GO TO 629
NS TR BG=ID B( 1 )+ !AR H( IPL USR)
IF(NSTRBG,GT.LB)GO TO 634
IDP=K-1
DO 635 IDD=l, !DP
63 5 IF(NSTRBG. EQ .• IDB(IDD))GO TO 634
IDB(K)=NSTR BG
K=K+l
634 I=I+l
GO TO 636
629 IPLUSR=IPLUSR+l
MM=K-1
I= 1
631 CONTINUE
IPLUSR=IR+l
632 KLA=LB-IFP(lPLUSR)
KSNR=KLA-KPLAN(KLA, IDB, K)
IARH(IR )=IFP(IPL USR )+KSNR
WRITE(6 , 922) IR, IARH(IR)
922 FOR MAT ( 1 X, 1IR= 1 , .:::6,' IA RH( IR )= ' , !6)
IR=IR-1
IF(IR, EQ. 0) RETURN l
RETURN
END

FUNCTION KPLAN(KLA, IDA, K)

DIMENSION IDA(300)
C FIND KPLAN< DRN
KPLAN=0
MM=K-1
DO 530 I=l, MM
IF(IDA(I). LT. KPLAN. OR. IDA(!). GT. KLA) GO TO 530
KPLAN=IDA(I)
53 0 CONTINUE
WRITE(6, 924) KPLAN
924 FORMAT(lX, 'KPLAN=',!6)
RETURN
END