LINEAR PROGRAMMING II 1

Learning Objective 1: Recognize problems that

linear programming can handle.
Linear programming lets you optimize an
objective function subject to some constraints.
The objective function and constraints are all
linear.
Learning objective 2: Know the elements of a linear programming problem -- what you need to
calculate a solution.
The elements are
(1) an objective function that shows the cost or profit depending on what choices you make,
(2) constraint inequalities that show the limits of what you can do, and
(3) non-negativity restrictions, because you cannot turn outputs back into inputs.
Linear Programming II: Minimization
2006 Samuel L. Baker
Assignment 11 is on page 16.
Introduction
A minimization problem minimizes the value of
the objective function rather than maximizing it.
Minimization problems generally involve finding
the least-cost way to meet a set of requirements.
Classic example -- feeding farm animals.
Animals need:
14 units of nutrient A,
12 units of nutrient B, and
18 units of nutrient C.
Two feed grains are available, X and Y.
A bag of X has 2 units of A, 1 unit of B, and 1 unit of C.
A bag of Y has 1 unit of A, 1 unit of B, and 3 units of C.
A bag of X costs $2. A bag of Y costs $4.
Minimize the cost of meeting the nutrient requirements.
To solve, express the problem in equation form:
Cost = 2X + 4Y objective function to be minimized
Constraints:
2X + 1Y $ 14 nutrient A requirement
1X + 1Y $ 12 nutrient B requirement
1X + 3Y $ 18 nutrient C requirement
8 8
Read vertically to see how much of each nutrient is in each grain.
X $ 0, Y $ 0 non-negativity
LINEAR PROGRAMMING II 2
Feasible region in minimization problem Isocost lines and optimum
Learning objective 3: Understand the principles that the computer uses to solve a linear programming
problem.
The computer uses the simplex method to systematically trace around the feasible area (the simplex) in
the figure above. It goes from one corner to the next until it finds the best value for the objective
function.
Learning objective 4a: What linear programming problems have no solution?
Those that have no feasible area. This means it's impossible to satisfy all the constraints at once.
Learning objective 4b: What difference does linearity make?
If the constraints are not linear, the feasible area has curved edges. The simplex method doesn't work
because you can't be sure that the solution is at a corner. The solution may be in the middle of a curved
edge. If the objective function is not linear, the solution may not even be on an edge. It may be in the
interior of the feasible area.
Graph method of solution
Graph the constraints as equalities, like before. The constraints are now $ rather than #, so the feasible
area is everything to the right and above all of the constraint lines. You want to find the lowest cost point
of this area. It will be a corner. You superimpose isocost lines that you get from the objective function.
When you find an isocost line that just nicks one corner of the area, that corner is the solution.
LINEAR PROGRAMMING II 3
The simplex method of solution for minimization problems
The simplex method gets more complicated when you have constraints that are >. The method can't start at
the origin (with X = 0, Y = 0, etc.) because that is not a feasible solution. To find a starting place, we
invent some new variables in addition to the slack variables. The new variables are called "artificial" and
are conventionally labelled with A's. Artificial variables allow us to automatically come up with a feasible
solution to start the simplex method. In this case, the starting feasible solution is A
1
=14, A
2
=12, A
3
=18,
and all the rest (X, Y, and the Ss) =0.
The equations now look like this.
2X + 1Y - 1S
1
+ 0S
2
+ 0S
3
+ 1A
1
+ 0A
2
+ 0A
3
= 14
1X + 1Y + 0S
1
- 1S
2
+ 0S
3
+ 0A
1
+ 1A
2
+ 0A
3
= 12
1X + 3Y + 0S
1
+ 0S
2
- 1S
3
+ 0A
1
+ 0A
2
+ 1A
3
= 18
Each constraint equation has one slack variable (S) and one artificial variable (A) with non-zero
coefficients. The slack variable in each equation has a -1 coefficient because in a minimization problem we
have to meet or exceed each constraint. This means we have to subtract the slack variable rather than add
it as we did when we were maximizing the objective function.
The objective function, which we want to minimize, is expanded to:
2X + 4Y + 0S
1
+ 0S
2
+ 0S
3
+ MA
1
+ MA
2
+ MA
3
= Cost
M stands for a mammoth number, meaning a very large unit cost for the A's, large enough to be sure that
none of the A's will be in the final solution.
In matrix form we have:
X Y S
1
S
2
S
3
A
1
A
2
A
3
2 1 -1 0 0 1 0 0 14
1 1 0 -1 0 0 1 0 12
1 3 0 0 -1 0 0 1 18
-2 -4 0 0 0 -M -M -M 0
The first thing we do is to get rid of those -M's that are at the bottom of the three columns for the A's. To
do this, multiply each of the constraint rows by M and add, element by element, to the last row.
2 1 -1 0 0 1 0 0 14
1 1 0 -1 0 0 1 0 12
1 3 0 0 -1 0 0 1 18
4M-2 5M-4 -M -M -M 0 0 0 44M
The columns for the A's now have one 1 and the rest 0's. The A's, and only them, are in the initial basic
feasible solution. The solution is A
1
= 14, A
2
= 12, A
3
= 18, all X's and S's = 0. Cost is 44M, a
monstrously mammoth number that we want to reduce. We then proceed with the simplex method as
before. The pivot column is the second, because it has the biggest number (5M-4) in the bottom row.
LINEAR PROGRAMMING II 4
2 1 -1 0 0 1 0 0 14
1 1 0 -1 0 0 1 0 12
1 3 0 0 -1 0 0 1 18
4M-2 5M-4 -M -M -M 0 0 0 44M
To find the pivot row, we divide each number in the last column by the corresponding number in the second
column, to get 14/1=14, 12/1=1, and 18/3=6. 6 is the smallest so the third row is the pivot row. The pivot
element is the 3.
The next step is to use that third row to eliminate the other elements in the second column. We multiply the
row by 1/3 and subtract that from the first row. We do the same for the second row. For the bottom row
we multiply the third row by (4/3 - 5M/3) and subtract that from the fourth row. For our new third row,
we divide the original third row by 3. We get:
X Y S
1
S
2
S
3
A
1
A
2
A
3
1.67 0 -1 0 0.33 1 0 -0.33 8
0.67 0 0 -1 0.33 0 1 -0.33 6
0.33 1 0 0 -0.33 0 0 0.33 6
2.33M-0.67 0 -M -M-1.33 0.67M 0 0 0 14M+24
This puts Y in the solution, along with A1 and A3. The next iteration adds X to the solution. The iteration
after that adds S
1
. The final tableau looks like this:
X Y S
1
S
2
S
3
A
1
A
2
A
3
1 0 0 -1.5 0.5 0 1.5 -0.5 9
0 0 1 -2.5 0.5 -1 2.5 -0.5 7
0 1 0 0.5 -0.5 0 -0.5 0.5 6
0 0 0 -1 -1 -M 1-M 1-M 30
This is the final tableau because all the numbers in the bottom row are 0 or negative. Notice that the only
M's left are under the A's.
The non-zero variables in the solution are shown by the columns that have one 1 and the rest 0's. The
solution amounts are in the right column, and match up with X, Y, and S
1
according to where the 1 in each
column is. The solution is: X = 9, Y = 3, S
1
= 7. S2 and S3 are not in the final solution, so S
2
= S
3
= 0.
The total cost is $30.
The shadow prices for the requirements are in the bottom row, in columns for S2 and S3. Both S
2
and S
3
have a shadow price of -$1. These mean that it would cost you $1 if you had to provide one more unit of
either nutrient B or nutrient C. Having to provide one more unit of nutrient A costs nothing because you
are already providing 7 extra units of A.
LINEAR PROGRAMMING II 5
Learning objective 5: Be able to solve small linear programming problems yourself.
The tricky part is setting up the objective function and the constraints. Write them on paper (without
the Ss and As). Then let your spreadsheet crunch them for you.
Minimizing with Excel
Fortunately, you dont have to do the simplex method by hand anymore. You can get Excel or Quattro Pro
to do it for you. The challenge becomes setting the problem up properly.
Here is the minimization example, as set up in Excel:
Column B is for Grain X. Column C
is for Grain Y. Column D has
formulas for total cost and total
amounts of nutrients provided. If you
type the formula in D3 first, and use
$ signs in the right places, you can
copy D3 to D6:D8.
Heres the same
spreadsheet, showing the
formulas in the D
column.
LINEAR PROGRAMMING II 6
Select Tools and then Solver... from Excels menu.
If you are re-using a
spreadsheet from an
earlier assignment, the
safest way to start is to
click on the Reset All
button. This clears all the
old settings and cell
choices, preventing some
elusive bugs.
Then set the Target Cell
to D3 and the Changing
Cells to B2:C2.
On the Equal To: line, check Min. If you dont do this, your solution will be unbounded. Solver will fly
off looking for the upper-right corner of your boundless feasible region.
Click on the Add button to add constraints.
The nutrient amounts our grain
provides, in D6:D8, must be greater
than or equal to the required nutrients
in E6:E8. Be sure to change the
relation in the center to >=.
Then click on this boxs Add button to
save this constraint and get ready to
add another one.
This time, its the non-negativity
constraints. B2 and C2 must be >= 0.
Click on OK once this box is filled in as
shown, because were done adding
constraints.
LINEAR PROGRAMMING II 7
Heres what the Solver
Parameters box should
look like when youre
ready to solve.
Click on Solve.
Done! A Solver Results
dialog box pops up in front of
your spreadsheet.
In the Reports section of the
Solver Results box, select the
Answer and Sensitivity
Reports. Then click OK.
You will then see that Excel has filled
in the best solution, as shown to the
right.
This says that we satisfy the
requirements for the least possible
cost with 9 bags of X and 3 bags of
Y. The cost is $30.
LINEAR PROGRAMMING II 8
Click on the Sensitivity Report tab. This time the
shadow prices are called Lagrange Multipliers.
Why not Shadow Prices? My bad! I did not
tell you to click the Solvers Options button and
check the Assume Linear box. It does not really
matter. The numbers are the same. Only the
name is different.
Anyway, the Lagrange Multipliers, which are the
Shadow Prices, tell us that if we had to provide
another unit of nutrient B, it would cost $1 more.
Providing another unit of C would also cost $1
more. In other words, if vitamin B or vitamin C
pills were available, we would buy them if the
price was less than $1 per unit. We would not buy any vitamin A pills, because the shadow price for A is
$0. The constraint for A is not binding.
The are two more shadow prices hiding in the Sensitivity Report. They are under the Reduced Gradient
heading above. In this example, they tell us how much the non-negativity constraints are costing us.
Which is nothing, because we would not want to turn nutrients into grain the way these numbers are. In
one of the homework problems, a Reduced Gradient number may be important.
Modifications and extensions
A. Multiple Optimal Solutions
If a constraint happens to be parallel to the isocost lines, there will be multiple solutions, if that constraint
turns out to be binding. I can convert this example into one that multiple solutions by making the prices of
both grains the same. This makes the isocost lines parallel to the nutrient B constraint.
Now both grains cost the
same, $3 per bag, and the
two grains have the same
amount of nutrient B.
When the cost per unit of a
nutrient is the same for two
grains, the isocost lines are
parallel to the constraint
for that nutrient.
The shading in this diagram is trying to show that the nutrient B constraint is parallel to the cost objective
function.
LINEAR PROGRAMMING II 9
One edge of the feasible region is now parallel to
the iso-cost lines. The diagram to the right shows
how the lowest iso-cost line that touches the
feasible region touches it along a whole line
segment, not just at one corner.
Excels solver does not directly tell you that there
are multiple solutions. Instead, it seems to pick
one solution arbitrarily. When I solved this with
Excel, I got an arbitrary value near the middle of
this line segment. This was a true solution, but it
was only one of many. Any combination of X and
Y that sum to 12 from (2,10) to (9,3) is an
optimum.
I solved again in Excel, this time first clicking the Options button and checking the Assume Linear Model
box. Excel came back with the (9,3) solution, which is at one end of the line segment of optimal solutions.
The senstivity
report, shown
here, does
have a subtle
indication that
something is
up: The
Shadow Price
for nutrient C
is $0, even
though there
is no slack in
nutrient C,
the amount of
C provided being equal to the amount required.
The $0 shadow price for C means that the C constraint is not really active at the (9,3) solution. Ordinarily
in a linear programming problem, the number of active constraints (constraints with positive shadow
prices) equals the number of variables. We have two variables, X and Y, so ordinarily we would have two
active constraints. The fact that we have only one active constraint indicates that we have parallelism, and
that the solution displayed is one among many.
LINEAR PROGRAMMING II 10
Adding constraint that x $ 5
B. Mixing $ and # constraints
Actually, we're already doing this with the non-
negativity constraints, but here's a further use of
the same technique.
Suppose in the desks and tables problem from last
week that we have an order for 5 desks that we are
obligated to fill. We have to produce at least 5
desks. The effect of adding this constraint is to cut
off part of the feasible area, from the left.
The new optimum turns out to be 5 desks and 5
tables. Profit is $35. Without the added
constraint, 4 desks and 6 tables were produced,
and the profit was $36. The added constraint
costs $1 in lost profit.
Doing this in Excel requires adding another constraint. The best way to do this is to add another constraint
row to the spreadsheet. Being obligated to supply 5 desks is expressed as a constraint that
1Desks + 0Tables >= 5. The 1 and 0 in that equation go in cells B9 and C9. Copy the formula in D8
and paste it to D9.
The added obligated constraint.
Then, in Solver, add the new constraint
that D9 >= E9. Be sure to get the
greater-than-or-equal-to sign showing in
the middle of the Add Constraint dialog
box.
Alternatively, you could just add a new
constraint that B2 >= 5. The problem
with this is that this new constraint is not visible unless you bring up the Solver dialog box. Six months
from now, you might come back to your spreadsheet and not remember that you added this constraint.
After Solving, the Sensitivity Report shows a shadow price for the obligated constraint of -1. This
means that every extra desk you are obligated to make reduces your profit by $1. This is consistent with
LINEAR PROGRAMMING II 11
the finding that adding the obligated constraint made the profit $35, instead of the $36 you made without
the new constraint. (Another advantage of putting this new constraint in the 9
th
row, instead of just adding
a constraint to Solver that B2 >= 5, is that the shadow price shows up properly labeled in the Sensitivity
Report.)
Synonyms for Shadow Price
Depending on how you which software you are using, whether you specify a linear model, and how you
handle constraints on single amounts (like the B2 >= 5 above), the what-its-worth-to-you-for-another-unit
prices may also be called:
< Shadow Prices
< Lagrange Multiplilers
< Opportunity Cost
< Reduced Cost
< Reduced Gradient
You may have look up and down the Sensitivity Report to find all the shadow prices in their various guises.
Applications
The computer takes care of the math, so the hard part is figuring out how to set the problem up.
A. Staff Scheduling
Suppose you have these weekly staffing requirements:
Day Staff
Needed
)))))))))))))))))))
Mon 180
Tue 160
Wed 150
Thu 160
Fri 190
Sat 140
Sun 120
Each employee works 5 days in a row, then he/she is off for two days. Employees may start on any day of
the week. Each employee costs $400 per week. The objective is to minimize the total cost of meeting the
requirements.
Here is one possible spreadsheet setup. It uses 0's and 1's to put in the constraints:
LINEAR PROGRAMMING II 12
To get the logic behind this spreadsheet, consider row 6,
for example. It's the constraint for Monday. The total
number of people on the job on Monday equals the
number of people who work Monday-Friday, plus
Thursday-Monday, Friday-Tuesday,
Saturday-Wednesday, and Sunday-Thursday. Row 6 has
1's in those columns. People who work Tuesday-Saturday
and Wednesday-Sunday are not working Monday, so Row
6 has 0's in those columns. Cell I6's SUMPRODUCT
formula calculates the total number of people working
Monday. It must be greater than or equal to what is in J6.
An alternative way to set up the problem is:
The formulas in Column E do what the first layout does with 0's and 1's. The idea behind them is, for
example, that the number of staff working Monday equals total number of people who started their 5-day
work week on Monday, Thursday, Friday, Saturday, and Sunday. The constraints are that each of the
formulas in the E column has to be greater than or equal to the corresponding Need amount in the F
LINEAR PROGRAMMING II 13
column.
Both methods also need the non-negativity constraints, that the cells containing the amounts to hire are all
>= 0.
B. Transportation
You are distributing pharmaceuticals from two locations. Location A has 100 units. Location B has 250
units. You have customers at locations C, D, and E, who want 50, 120, and 180 units respectively.
Minimize cost, based on these unit transportation costs:
From: A B
To:
C $5 $3
D $10 $4
E $12 $9
Here is a spreadsheet layout for this transportation problem:
The key insight is that the variables are the six possible routes, that is, the six possible combinations of the
two sources and the three destinations.
The constraints are a mixture of >= and <=. Be careful when you enter these to be sure that they are all in
the right directions. The constraints are:
H5:H6 <= I5:I6
H7:H9 >= I7:I9
B2:G2 >= 0
If Solver gives you an error message that the solution is unbounded, check the directions of your > and <
signs. Also check that you did include the non-negativity constraint, that the amounts are all >= 0.
LINEAR PROGRAMMING II 14
The dual illustrated
Every linear programming problem has a "dual" -- a restatement of the problem that switches the roles of
the constraints and the activities. The original problem is called the "primal."
Here is the desks and tables example, showing the primal and the dual. The numbers are the same in the
primal and the dual, but the columns in one are the rows in the other, the constraint constants in one are the
coefficients of the objective function in the other, and the directions of the inequalities are switched. In the
primal version of the desks and tables problem we maximize the objective function. We minimize the
objective function in the dual.
Primal:
The constraints in the primal are:
D6:D8 <= E6:E8
B2:C2 >= 0.
The objective is to maximize D3.
Dual:
The constraints in the dual are:
E5:E6 >= F5:F6
B2:D2 >= 0.
The objective is to minimize E3.
In the dual you are doing some accounting. You are allocating the profit value of each finished product
(desk or table) to the inputs used to make it. These inputs' accounting values tell you how much you should
be willing to pay if you can buy more inputs. These accounting values are the shadow prices of the inputs.
Because they come from the Dual, they are sometimes called dual values.
The E5 >= F5 constraint in the dual says that you must choose values for the three inputs such that the
total value of the inputs into a desk is at least as great as $3, the value of a desk. The E6 >= F6 constraint
in the dual says that the values you put on those three inputs must also be such that the total value of the
inputs that go into a table is at least as big as $4, the value of a table. The dual's objective function says
that you want to chose values that minimize the total value of your initial stock of available resources. A
mind-bender, isn't it?
The solutions are on the next page.
LINEAR PROGRAMMING II 15
The primals Solver Sensitivity Report
The duals Solver Sensitivity Report
The primals solution:
The duals solution:
The primals solution amounts (4 and 6) are the shadow
prices in the dual. Likewise, the duals solution values
(0, .667, and 1) are the shadow prices in the primal.
The simplex method solves the primal and the dual
simultaneously. This is how Excel gets the numbers it
puts in the Sensitivity Report.
Your turn:
LINEAR PROGRAMMING II 16
Assignment 11
Write out your answers in sentences. Show your work. Write out your equations for the
constraints and objective function. Include a printout of your spreadsheet results or attach a copy of
your spreadsheet file.
1. Use the graph method and then the computer to solve this one:
You are the hospital gardener. The flowers require 28 pounds of nitrogen, 30 pounds of
phosphates, and 4 pounds of potash. You have two fertilizers to use. One kind costs $4 per bag. It
contains 4 pounds of nitrogen, 2 pounds of phosphates, and no potash per bag. The other kind costs $5
per bag. It contains 2 pounds of nitrogen, 3 pounds of phosphates, and 1 pound of potash.
a. How many bags of each kind of fertilizer do you use to minimize your costs?
b. The CEO's brother owns the company that makes the first kind of fertilizer. You must buy at
least 6 bags of it. Now how much of each kind of fertilizer do you use? (Be sure to click Solvers
Options button and check Assume Linear Model.)
c. For b, what do the shadow prices mean? Explain all of them, including the shadow price for
the constraint about having to buy 6 from the CEOs brother. (It may be called Shadow Price, Lagrange
Multiplier, Reduced Cost, or Reduced Gradient, depending on how you set up your spreadsheet.)
2. You have to schedule aides to cover the following daily requirements:
Monday 80
Tuesday 85
Wednesday 75
Thursday 80
Friday 90
Saturday 70
Sunday 60
You can start an aide on any day of the week. An aide works 5 consecutive days, then gets 2 days
off. Aides cost $40 per day on weekdays, $44 on Saturday and Sunday. This means that a work week
that includes no weekend days costs $200. A work week that includes one weekend day costs $204. A
work week that includes both weekend days costs $208. How do you hire aides so as to minimize your
total cost per week? What will your total cost per week be?
3. Let's build on the transportation problem example on page 13. Imagine that more problems have
arisen. Depot A and Customers C and D are in Canada. Depot B and Customer E are in the USA.
Canada will not allow you to import more than 100 units from the USA. Also, there's a weak bridge
along the road from A to C. No more than 40 units can be sent that way. How much do you move from
each depot to each customer to meet all the requirements at minimum cost? If you could pay someone to
smuggle units in excess of 100 into Canada, how much would it be worth to you per unit?