Solving Linear Programs 1.0 Basic Form: C) X (H X F

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Solving Linear Programs

1.0 Basic form

We go back to the form of the basic optimization


problem we have considered, which is
min f ( x)
s.t. h( x )  c
(1)
g ( x)b
We recall that a linear program (LP) requires all
functions f, h, and g to be linear in the variables x.

We make four comments regarding (1).


1. Equality constraints: The equality constraints may
be converted into two inequality constraints via the
following approach:
h( x )  c
h( x )  c 
h( x )  c (2)
and we may then reverse the sign of the second
inequality, resulting in:
h( x )  c
h( x )  c 
 h( x )  c (3)
This means we may include all of our equality
constraints h(x)=c from our general form (1) in our
inequality constraints g(x)< b, so that the general
form of our problem becomes:

1
min f ( x)
s.t. g ( x)b (4)

2. Nonnegativity constraints: The form given in (4)


places no restriction on the sign of the decision
variables in x. Most problems require all decision
variables to be nonnegative. For example,
generation offers and demand bids are typically
this way. This is convenient, because the algorithm
we will present to solve LPs requires
nonnegativity on the decision variables. It is the
case, however, that some problems need to allow
negativity for some or all decision variables. For
example, we might like to develop a generation
dispatch function that computes changes in
generation rather than generation. For such cases,
it is possible to convert one decision variable
without the nonnegativity constraint into one, or
two decision variables with nonnegativity
constraints. We will not go into that here, but
suffice it to say there are many references that
describe how to do this, including [1, pg. 83-86].
Given this, we pose our general form of the
optimization as

2
min f ( x)
s.t. g ( x)b
(5)
x0
3. No lower bounds on g(x): It is possible that a
problem has constraints like cl<g(x)<ch. In this
case, the right hand side g(x)<ch is already in the
correct form, but the left hand side cl<g(x) is not.
In this case, we can multiply both sides by -1,
resulting in
g(x)<-cl (6)
The negative right-hand-side is addressed in the
next bullet-point.
4. Negative right-hand-sides: We also require that all
elements of b be nonnegative. This may seem to
contradict the step we took in eq. (3) and (6)
above. However, we will see at the end of these
notes that it is possible to convert inequalities like
these, with a negative right-hand-side, to the
desirable form.

2.0 A simple solution approach

We concluded our last set of notes (IntroLP) with the


statement

3
“We will call such intersection points corner
points. Therefore we see that our solution will
always be at a corner point. This provides us
with a basis for solution to LPs: Search the
corner points!”
This is an effective approach, and if you take it, you
will always find the right answer. However, you may
also be doing a great deal of work. In our second
example (Section 3.0) of our previous notes
(IntroLP), we considered following LP:
max f ( x, y)  5x  8 y
Subject to
40 x  30 y  480 (person 1)
24 x  32 y  480 (person 2)
20 x  24 y  480 (person 3)
x  0, y  0
and the constraints are visualized in Fig. 1.

4
Fig. 1: Constraints for example problem

Here we see there are four corner points to search:


(0,0), (12,0), (1.7143, 13.7143), and (0,15). However,
we must make an important distinction here. These
are feasible corner points. Because we know the
solution must be feasible, these are the right points to
search. Yet there are other infeasible corner points.
For example, (0,16), (0,20), (20,0), and (24,0) are
four such infeasible corner points. And there are two
more that are outside the plane that we have plotted
in Fig. 1. Counting them up, we see there are a total
of 10 corner points, one for every pair of constraints.

5
This number of corner points is a combinatorial
problem, characterized by 5 distinct things
(constraints) taken 2 at a time, i.e.,
5  5! 5 * 4 * 3 * 2 *1
2 2!(5  2)! (2 *1)(3 * 2 *1)  10
 
 
We can easily distinguish feasible and infeasible
corner points from Fig. 1. However, it will not be so
easy for larger problems, especially when there are
many decision variables and we cannot easily
visualize the constraints in 2-D as we are doing here.
One could perhaps devise a means to check them all,
but it would be highly computational. For example,
consider having just 40 constraints, there would be
780 corner points to check. Some problems have
millions of constraints.

3.0 A better solution approach

So we need to develop a more effective strategy. To


do so, let’s consider a graphical portrayal of some
LP, as shown in Fig. 2.

In Fig. 2, the dashed lines are the constraint


boundaries, and the thick solid blue lines enclose the
feasible region. The thin solid lines show the
contours of constant objective function.

6
5
4 2
2 6
2

100
90
3
2 7 80
2
70
60
50
8
2 40
2
30
9
20
1 10
2 5 10

Fig. 2: Feasible region and contours for some LP

It is easy to see in Fig. 2 that if we are minimizing,


the solution is corner point 1, and the minimum value
is 5.

Likewise, if we are maximizing, the solution is


corner point 5, and the maximum value is 100.

We consider a strategy for solving this problem. This


strategy depends on the following two definitions:
 Adjacent corner points are connected by a single

line segment on the boundary of the feasible region.


 One corner point is better than another if it has a

higher value of the objective function f.

7
Our strategy is as follows:
1. Pick a corner point at random.
2. Move to an adjacent corner point that is better.
a. If there are two that are better, move to the one
that is best.
b. If there are no better adjacent corner points,
the current corner point is the solution to the
problem.
Let’s apply this strategy to the problem of Fig. 2,
assuming we are maximizing the function. We also
assume that we initially choose corner point 1.

From corner point 1, we can either move to corner


point 2 or 11. But the objective function value at
corner point 2, f2=20, whereas the objective function
value at corner point 11 is only f11=10. Although both
are better than f1=5, we choose to move to corner
point 2 since it is better.
From corner point 2, f2=20, we can move to corner
point 1, f1=5 or corner point 3, f3=60. Corner point 1
is not an option since it does not get better. But
corner point 3 is better, with f3=60, so we move there.

In like fashion, we move to corner point 4, f4=95, and


then to corner point 5, f5=100.

8
At corner point 5, there are two options: corner point
4, with f4=95, or corner point 6, with f6=97. Both of
these are worse than corner point 5. So we are done,
and corner point 5 is the solution with f5=100 as the
maximum value of the problem.

From this example, we may conjecture a condition


for optimality:
If a corner point feasible solution is equal to or
better than all its adjacent corner point feasible
solutions, then it is equal to or better than all
other corner point feasible solutions, i.e., it is
optimal.
Formal proofs of this optimality condition are
available in some texts; here, we simply state the
essence of such proofs, which is contained in the
following two points.
1. If the objective function monotonically increases
(decreases) in some direction within the decision-
vector space, then each adjacent corner point will
become progressively better in the direction of
objective function increase (decrease) such that the
last corner point must have two adjacent corner
points that are worse.
2. The monotonicity of objective function increase
(decrease) is guaranteed by its linearity.

9
With the above optimality condition in place, we may
outline the algorithm that we are going to study for
solving linear programs. It is called the Simplex
Method, and at a high level, is like this [1]:
1. Initialization: Start at a corner point solution.
2. Iterative step: Move to a better adjacent corner
point feasible solution.
3. Optimality test: Determine if the current feasible
corner point is optimal using our optimality test (if
none of its adjacent feasible corner points are
better, then the current feasible corner point is
optimal).
a. If the current feasible corner point is
optimal, the solution has been found, and the
method terminates.
b. If the current feasible corner point is not
optimal, then go to 2.

4.0 A word about the simplex method

The simplex method was developed in 1947 by


George Dantzig (1914-2005) who worked for the US
Air Force in the Pentegon to find better ways to plan
the Air Force activities. He was trained as a
mathematician but had significant experience in
developing the plans that the air force required.

10
It is hard to conceptualize now, but at that time, there
was no notion of an objective function. Neither was
there any understanding that physical constraints on
resources could be represented by linear inequalities.
Dantzig recognized both of these; in addition, he
developed the simplex method we are about to study.

The simplex method almost immediately


revolutionized many fields, among which were
planning, production, and economics.

It is interesting to note what brought this


development to fruition:
 A war (and the needs of the Air Force)

 A person trained in mathematics and with


significant practical experience in solving the
problems at hand who, it seems, needed a job.

I have posted on our website a short and very


readable summary paper written by George Dantzig
in 2002 on how the simplex method came to be.

11
5.0 Setting up the simplex Method

The material of this section is adapted from [1].

The simplex method is comprised of a number of


algebraic manipulations. These manipulations are
made much easier if we first convert the inequality
constraints into equality constraints by introducing
slack variables.

We use another simple example to explain this idea.


The example is as follows:
max F  3 x1  5 x2
s.t. x1 4
2 x2  12
3 x1  2 x2  18
x1  0, x2  0
Consider the first constraint x1<4. The slack variable
for this constraint is
x3=4-x1 (6)
which represents the “slack” between the two sides of
the inequality x1<4. If the “slack” is 0, then the
inequality is satisfied with equality, and there is
really “no slack.” This variable cannot be negative,
otherwise, x1>4.

12
Therefore, the first inequality, x1<4, may be replaced
with
x1  x3  4, x3  0 (7)
We may similarly introduce slack variables into the
other constraints, so that our original LP is converted
to the following equivalent LP.
max F  3 x1  5 x2
s.t.
x1  x3  4
2 x2  x4  12
3 x1  2 x2  x5  18
(8)

x1  0, x2  0, x3  0, x4  0, x5  0
We need a few definitions:
 Equality form: In contrast to the original inequality

form, the equality form of the problem has all


inequality constraints converted to equality
constraints via introduction of slack variables.
 Augmented solution: A solution to the LP that

includes appropriate values of the slack variables (in


addition to the values of the decision variables). For
example, a solution to the original LP may be stated
as (x1,x2)=(3,2) whereas an augmented solution would
be stated as (x1,x2, x3, x4, x5)=(3,2,1,8,5).

13
 Basic solution: An augmented corner-point solution.
For example, consider the corner point infeasible
solution (4,6) in the example. Augmenting it with the
slack variable values x3=0, x4=0, and x5=-6 yields the
corresponding basic solution (4,6,0,0,-6). This basic
solution is infeasible, as indicated by the presence of
the negative slack variable x5. This point is illustrated
by the ‘O’ in Fig. 3.
 Basic feasible solution: A feasible augmented
corner-point solution. For example, consider the
corner point feasible solution (0,6) in the example.
Augmenting it with the slack variables x3=4, x4=0,
and x5=6 yields the corresponding basic feasible
solution (0,6,4,0,6). This basic solution is feasible.
This point is illustrated by the ‘X’ in Fig. 3.

14
Fig. 3: Illustration

6.0 Final comment on general form of problem

Recall our general LP form:


min f ( x)
s.t. g ( x)b
(5)
x0
We mentioned in Section 1.0 of these notes, under
point 3, that we required all elements of b to be
nonnegative. Yet, in our point #1, we indicated we
could handle equality constraints via the following
transformation:

15
h( x )  c
h( x )  c 
 h( x )  c (3)
This results in a negative right-hand-side;
assuming elements of c are all positive, then the
second equation above would have all negative
right-hand sides.

The way to handle this is to first convert the


problem into equality form via introduction of
slack variables. Then another slack variable can be
added for all equations having negative right-hand-
sides. An example will clarify.

Let’s assume that our last statement of our


example problems has one of the constraints with a
negative right-hand-side, per below (the last
equation, with -18 as the negative right-hand-side).

max F  3 x1  5 x2
s.t.
x1  x3  4
2 x2  x4  12
3 x1  2 x2  x5   18

x1  0, x2  0, x3  0, x4  0, x5  0

16
Let’s extract the one equation with the negative
right-hand-side:
3x1  2 x2  x5  18 (9)
Now we multiply by -1 to get
 3x1  2 x2  x5  18 (10)
Although this creates the positive right-hand-side
that we need, we will see later on that our
initialization procedure to find a feasible solution
will fail here, because it will result in x5=-18 and
therefore violates variable nonnegativity. As a
result, we must add another slack variable here,
resulting in
 3x1  2 x2  x5  x6  18 (10)
Now we have that the right-hand-side is positive
and our initialization procedure will result in
x6=18, satisfying decision variable nonnegativity
and nonnegativity on the element in b.

[1] F. Hillier and G. Lieberman, “Introduction to Operations Research,” 4th edition,


Holden-Day, Oakland California, 1986.

17

You might also like