Integer Linear Optimization

Models 1
 Learning Objectives
• Differentiate types of integer linear optimization models –
all-integer, mixed-integer and binary integer.
• Apply geometric approach in solving linear all-integer
optimization.
• Use Excel Solver in integer linear optimization problems.
• Solve integer linear optimization problems involving binary
variables.
• Generate alternatives in binary optimization.
2
Introduction

3
 Introduction
• Integer linear programs: Problems that are modeled as linear
programs with the additional requirement that one or more
variables must be integer.

• The objective is to provide an applications-oriented introduction

to integer linear programming.

4
Types of Integer Linear
Optimization Models

5
 Types of Integer Linear Optimization Models
• All-integer linear program: If all variables are required to be integer.

• Mixed-integer linear program: If some, but not necessarily all, variables

are required to be integer.

• Binary integer linear program: The integer variables may take on only
the values 0 or 1.

LP Relaxation (linear programming relaxation) of the integer linear

program: The linear program that results from dropping the integer
requirements.

6
Eastborne Realty, An Example
of Integer Optimization

7
 Eastborne Realty, An Example of Integer
Optimization
• Eastborne Realty has $2 million available for the purchase of new
rental property.
• After an initial screening, Eastborne reduced the investment
alternatives to townhouses and apartment buildings.
• Each townhouse can be purchased for $282,000.
• Five are available.
• Each apartment building can be purchased for $400,000.
• The developer will construct as many buildings as Eastborne wants
to purchase.

8
 Eastborne Realty, An Example of Integer
Optimization

• Eastborne’s property manager can devote up to 140 hours per

month to these new properties.

• Each townhouse is expected to require 4 hours per month.

• Each apartment building is expected to require 40 hours per

month.

9
 Eastborne Realty, An Example of Integer
Optimization

• Annual cash flow estimated to be:

• $10,000 per townhouse.

• $15,000 per apartment building.

• After deducting mortgage payments and operating expenses.

• Eastborne’s owner would like to determine the number of

townhouses and the number of apartment buildings to purchase
to maximize annual cash flow.
10
 Eastborne Realty, An Example of Integer
Optimization
• Defining the decision variables:
• T = number of townhouses
• A = number of apartment buildings
• Objective function for cash flow (in thousands of dollars):
• Max 10T + 15A
• Three constraints must be satisfied:
• 282T + 400A ≤ 2000 Funds available ($1000s)
• 4T + 40A ≤ 140 Manager’s time (hours)
• T ≤ 5 Townhouses available
11
 Eastborne Realty, An Example of Integer
Optimization

• The variables T and A must be nonnegative.

• In addition, the purchase of a fractional number of townhouses

and/or a fractional number of apartment buildings is
unacceptable.

• T and A must be integer.

12
 Eastborne Realty, An Example of Integer
Optimization
• Model for the Eastborne Realty problem:
• Max 10T + 15A
• s.t.
282T + 400A ≤ 2000
4T + 40A ≤ 140
T ≤ 5
T, A ≥ 0 and integer

• This is a linear all-integer program.

13
Figure 1 - The Geometry of the Eastborne Realty
Problem

14
The Geometry of Linear All-Integer Optimization
• The shaded region is the feasible region of the LP Relaxation.
• The optimal linear programming solution is:
• Point b: The intersection of the Managers Time constraint and the
Available Funds constraint.
• T = 2.479 townhouses and A = 3.252 apartment buildings.
• The optimal value of the objective function is 73.574.
• Indicates an annual cash flow of $73,574.
• Unfortunately, Eastborne cannot purchase fractional numbers of
townhouses and apartment buildings.
• Further analysis is necessary.
15
The Geometry of Linear All-Integer Optimization
• In many cases, a noninteger solution can be rounded to obtain an
acceptable integer solution.
• For instance, a linear programming solution to a production scheduling
problem might call for the production of 15,132.4 cases of breakfast
cereal and the rounded integer solution of 15,132 cases would probably
have minimal impact on the value of the objective function and the
feasibility of the solution.
• A near-optimal solution is fine.
• However, rounding may not always be a good strategy.
• When the decision variables take on small values that have a major
impact on the value of the objective function or feasibility, an optimal
integer solution is needed.
16
The Geometry of Linear All-Integer Optimization

• Optimal solution to the LP Relaxation for Eastborne Realty:

• T = 2.479 townhouses and A = 3.252 apartment buildings.

• Rounding to an integer solution can be expected to have a

substantial economic impact on the problem:

• Because each townhouse costs $282,000 and each apartment

building costs $400,000.

17
The Geometry of Linear All-Integer Optimization

• Suppose, round the solution to the LP Relaxation:

• Integer solution: T = 2 and A = 3.

• Objective function value: 10(2) + 15(3) = 65.

• The annual cash flow of $65,000 is substantially less than the

annual cash flow of $73,574 provided by the solution to the LP
Relaxation.

18
The Geometry of Linear All-Integer Optimization

• Do other rounding possibilities exist?

• Exploring other rounding alternatives:

• Integer solution: T = 3 and A = 3.

• Objective function value: $282,000(3) + $400,000(3) = $3,738,000.

• Infeasible.

• Because $3,738,000 > $2,000,000, the available funds with Eastborne.

19
The Geometry of Linear All-Integer Optimization

• The rounded solution of T = 2 and A = 4 is also infeasible for the

same reason.

• At this point, the best feasible integer solution to the problem:

• Rounding to two townhouses and three apartment buildings.

• With an annual cash flow of $65,000.

• Unfortunately, it is not known if this solution is the best integer

solution to the problem.
20
The Geometry of Linear All-Integer Optimization

• Rounding to an integer solution is a trial-and-error approach.

• Each rounded solution must be evaluated for feasibility as well as

for its impact on the value of the objective function.

• Even when a rounded solution is feasible, there is no guarantee

that the optimal integer solution has been found.

21
The Geometry of Linear All-Integer Optimization

• What is the true feasible region for the Eastborne Realty

problem?

• As shown in Figure 9.1, the feasible region is the set of integer

points that lie within the feasible region of the LP Relaxation.

• There are 20 such feasible solutions (designated by blue dots in

the figure).

• The region bounded by the dashed lines is known as the convex

hull of the set of feasible integer solutions.
22
Solving Integer Optimization
Problems with Excel Solver

23
 Solving Integer Optimization Problems with
Excel Solver

• The worksheet formulation and solution for integer linear

programs is similar to that for linear programming problems.

• But some additional information must be provided when setting

up the Solver Parameters and Options dialog boxes.

• Constraints must be added in the Solver Parameters dialog box to

identify the integer variables.
• In addition, the value for Tolerance in the Integer Options dialog
box may need to be adjusted to obtain a solution.
24
Figure 2 - Eastborne Realty Spreadsheet Model

25
Figure 2 - Eastborne Realty Spreadsheet Model

26
Figure 3 - Solver Parameters Dialog Box for
Eastborne Realty

27
Figure 4 - Solver Options Dialog Box

28
Figure 5 - Excel Solver Answer Report for the
Eastborne Realty Problem

29
 Solving Integer Optimization Problems with
Excel Solver

• A cautionary note about sensitivity analysis

• Sensitivity analysis often is more crucial for integer linear

programming problems than for linear programming problems.

• A small change in one of the coefficients in the constraints can

cause a relatively large change in the value of the optimal
solution.

30
Applications Involving Binary
Variables

31
 Applications Involving Binary Variables
• Capital budgeting problem: A binary integer programming
problem that involves choosing which possible projects or
activities provide the best investment return.

• In a capital budgeting problem, the objective function is to

maximize the net present value of the capital budgeting projects.

32
 Applications Involving Binary Variables
• Illustration

• The Ice-Cold Refrigerator Company is considering investing in

several projects that have varying capital requirements over the
next four years.

• Faced with limited capital each year, management would like to

select the most profitable projects that it can afford.

33
Table 1 - Project Net Present Value, Capital Requirements, and
Available Capital for the Ice-Cold Refrigerator Company

34
 Applications Involving Binary Variables
• Let us define four binary decision variables:

• P = 1 if the plant expansion project is accepted; 0 if rejected

• W = 1 if the warehouse expansion project is accepted; 0 if rejected

• M = 1 if the new machinery project is accepted; 0 if rejected

• R = 1 if the new product research project is accepted; 0 if rejected

• This problem has four constraints:

• One for the funds available in each of the next four years.
35
 Applications Involving Binary Variables
• A binary integer linear programming model (thousands of
dollars):
• Max 90P + 40W + 10M + 37R
• s.t.
• 15P + 10W + 10M + 15R ≤ 40 (Year 1 capital available)
• 20P + 15W + 10R ≤ 50 (Year 2 capital available)
• 20P + 20W + 10R ≤ 40 (Year 3 capital available)
• 15P + 5W + 4M + 10R ≤ 35 (Year 4 capital available)
• P, W, M, R = 0, 1
36
Figure 6 - Ice-Cold Spreadsheet Model and Solver
Dialog Box

37
Figure 7 - Answer Report for Ice-Cold
Refrigerator

38
 Applications Involving Binary Variables
• In many applications, the cost of production has two components:

• a setup cost, which is a fixed cost.

• a variable cost, which is directly related to the production quantity.

• Fixed-cost problem: A binary mixed-integer programming

problem in which the binary variables represent whether an
activity, such as a production run, is undertaken (variable = 1) or
not (variable = 0).

39
 Applications Involving Binary Variables
• Illustration
• Consider the production problem faced by RMC Inc.
• Three raw materials are used to produce three products:
• a fuel additive
• a solvent base
• a carpet cleaning fluid
• Decision variables used are:
• F = tons of fuel additive produced
• S = tons of solvent base produced
• C = tons of carpet cleaning fluid produced
40
 Applications Involving Binary Variables
• A linear programming model of the RMC problem:

• Max 40F + 30S + 50C

• s.t.

• 0.4F + 0.5S + 0.6C ≤ 20 Material 1

• 0.2S + 0.1C ≤ 5 Material 2

• 0.6F + 0.3S + 0.3C ≤ 21 Material 3

• F, S, C ≥ 0
41
 Applications Involving Binary Variables
• Suppose that the RMC data are available concerning the setup cost and
the maximum production quantity for each of the three products.

• The modeling flexibility provided by binary variables can now be used to

incorporate the fixed setup costs into the production model.

• The binary variables are defined as:

• SF = 1 if the fuel additive is produced; 0 if not

• SS = 1 if the solvent base is produced; 0 if not

• SC = 1 if the carpet cleaning fluid is produced; 0 if not

42
 Applications Involving Binary Variables
• The fixed-cost model for the RMC problem with setups:
• Max 40F + 30S + 50C – 200SF – 50SS – 400SC
• s.t.
• 0.4F + 0.5S + 0.6C ≤ 20 Material 1
• 0.2S + 0.1C ≤ 5 Material 2
• 0.6F + 0.3S + 0.3C ≤ 21 Material 3
• F ≤ 50SF Maximum Fuel Additive
• S ≤ 25SS Maximum Solvent Base
• C ≤ 40SC Maximum Carpet Cleaning
• F, S, C ≥ 0; SF, SS, SC = 0 or 1 Refer to Figure 9.8 after this.
43
Figure 9 - Answer Report for RMC Production
Problem

44
 Applications Involving Binary Variables
• Location problem: A binary integer programming problem in
which the objective is to select the best locations to meet a stated
objective.

• Variations of this problem are known as covering problems.

45
 Applications Involving Binary Variables
• Illustration

• The long-range planning department for the Ohio Trust Company is considering
expanding its operation into a 20-county region in northeastern Ohio.

• Currently, Ohio Trust does not have a principal place of business in any of the
20 counties.

• To establish a new principal place of business, Ohio Trust must either obtain
approval for a new bank from the state’s superintendent of banks or purchase
an existing bank.
• Refer to Figure 9.10 here.
46
Figure 11 – Ohio Trust County Map

47
Table 2 - Counties in the Ohio Trust Expansion
Region

48
 Applications Involving Binary Variables
• As an initial step in its planning, Ohio Trust would like to
determine the minimum number of PPBs necessary to do
business throughout the 20-county region.

• A binary integer programming model can be used to solve this

location problem for Ohio Trust.

• Define the variables:

• xi = 1 if a PPB is established in county i; 0 otherwise

49
 Applications Involving Binary Variables
• The complete statement of the bank location problem:

50
Figure 11 - Optimal Solution to the Ohio Trust
Location Problem

51
 Applications Involving Binary Variables
• Conjoint analysis: A market research technique that can be used to learn
how prospective buyers of a product value the product’s attributes.

• Product design and market share optimization problem: Sometimes called

the share of choice problem, the choice of a product design that maximizes
the number of consumers preferring it.

• The results of conjoint analysis can be used in an integer programming

model of a product design and market share optimization problem.

• Part-worth: Utility value that a consumer attaches to each level of each

attribute.
52
 Applications Involving Binary Variables
• Illustration

• Salem Foods is planning to enter the frozen pizza market.

• Currently, two existing brands, Antonio’s and King’s, have the major
share of the market.

• In trying to develop a sausage pizza that will capture a substantial share

of the market, Salem determined that the four most important
attributes when consumers purchase a frozen sausage pizza are crust,
cheese, sauce, and sausage flavor. (Contd.)

53
 Applications Involving Binary Variables
• Crust: thin and thick
• Cheese: mozzarella and blend
• Sauce: smooth and chunky
• Sausage flavor: mild, medium, and hot

• In a typical conjoint analysis, a sample of consumers is asked to express

their preference for a product with chosen levels for the attributes.

• Then regression analysis is used to determine the part-worth for

each of the attribute levels.

54
 Table 3 - Part-worths for the Salem Foods
Problem

• The part-worths can be used to determine the overall value (utility) that
each consumer attaches to a particular type of pizza.
• In general, each consumer’s utility for a particular type of pizza is the
sum of the part-worths for the attributes of that type of pizza.
55
 Applications Involving Binary Variables
• Salem must design a pizza (choose the type of crust, cheese, sauce, and
sausage flavor) that will have the highest utility for a sufficient number
of people to ensure sufficient sales to justify making the product.
• Assuming the sample of eight consumers in the current study is
representative of the marketplace for frozen sausage pizza:
• Formulate and solve an integer programming model that can help
Salem come up with such a design. In marketing literature, the
problem being solved is called the share of choice problem.
• The decision variables are defined as:
• lij = 1 if Salem chooses level i for attribute j; 0 otherwise
• yk = 1 if consumer k chooses the Salem brand; 0 otherwise
56
 Applications Involving Binary Variables
• Model for the Salem pizza problem:
• Max y1 + y2 + . . . + y8
• s.t.

57
 Applications Involving Binary Variables
• The optimal solution to this 17-variable, 12-constraint integer linear program:

• l11 = l22 = l23 = l14 = 1 and y2 = y5 = y6 = y7 = 1.

• The value of the optimal solution is 4.

• Indicating that if Salem makes this type of pizza, it will be preferable to
the current favorite for four of the eight consumers.
• With l11 = l22 = l23 = l14 = 1, the pizza design that obtains the largest
market share for Salem has a thin crust, a cheese blend, a chunky sauce,
and mild-flavored sausage.
• Note also that with y2 = y5 = y6 = y7 = 1, consumers 2, 5, 6, and 7 will
prefer the Salem pizza.

58
Modeling Flexibility Provided
by Binary Variables

59
Modeling Flexibility Provided by Binary Variables

• Binary integer variables can be used to model:

• Multiple-choice and mutually exclusive constraints.

• Situations in which k projects out of a set of n projects must

be selected.

• Situations in which the acceptance of one project is

conditional on the acceptance of another project.

60
 Multiple-Choice and Mutually Exclusive
Constraints
• Multiple-choice constraint: A constraint requiring that the sum of two
or more binary variables equals one.

• Thus, any feasible solution makes a choice of which variable to set equal
to one.

• Mutually exclusive constraint: A constraint requiring that the sum of

two or more binary variables be less than or equal to one.

• Thus, if one of the variables equals one, the others must equal zero.

• However, all variables could equal zero.

61
 Multiple-Choice and Mutually Exclusive
Constraints
• Ice-Cold Refrigerator capital budgeting problem:

• W1 = 1 if the original warehouse expansion project is accepted; 0 if rejected

• W2 = 1 if the second warehouse expansion project is accepted; 0 if rejected

• W3 = 1 if the third warehouse expansion project is accepted; 0 if rejected

• The multiple-choice constraint reflecting the requirement that exactly

one of these projects must be selected:

• W1 + W 2 + W 3 = 1

62
 Multiple-Choice and Mutually Exclusive
Constraints
• Ice-Cold Refrigerator capital budgeting problem:

• If the requirement that one warehouse must be expanded did not exist,
the multiple-choice constraint could be modified as:

• W1 + W 2 + W3 = 1

• It allows for the case of no warehouse expansion: W1 = W2 = W3 = 0.

• It does not permit more than one warehouse to be expanded.

• This resulted in a mutually exclusive constraint.

63
 k Out of n Alternatives Constraint
• k out of n alternatives constraint: A constraint that requires the sum of
n binary variables to equal k.
• An extension of the multiple-choice constraint.
• In the capital budgeting problem, suppose that W1, W2, W3, W4, and W5
represent five potential warehouse expansion projects and that two of
the five projects must be accepted.
• Constraint satisfying this new requirement:
• W1 + W 2 + W3 + W 4 + W5 = 2
• If no more than two of the projects are to be selected:
• W1 + W 2 + W3 + W 4 + W5 ≤ 2

64
 Conditional and Corequisite Constraints
• Conditional constraint: A constraint involving binary variables that does not
allow certain variables to equal one unless certain other variables are equal to
one.
• In the capital budgeting problem, suppose that the warehouse expansion
project was conditional on the plant expansion project.
• With P representing plant expansion and W representing warehouse
expansion:
• Conditional constraint: W ≤ P
• P and W must each be 0 or 1.
• When P is 0, W will be forced to 0.
• When P is 1, W is allowed to be 1.
• Both the plant and the warehouse can be expanded. 65
 Conditional and Corequisite Constraints
• Corequisite constraint: A constraint requiring that two binary variables be
equal and that thus are both either in or out of the solution together.

• In the capital budgeting problem, if the warehouse expansion project had to be

accepted whenever the plant expansion project was, and vice versa:

• P and W represented corequisite constraint projects.

• To model such a situation, write the preceding constraint as an equality:

• W=P

• The constraint forces P and W to take on the same value.

66
Generating Alternatives in
Binary Optimization

67
 Generating Alternatives in Binary Optimization
• If alternative optimal solutions exist, it would be good for management to
know this:
• Because some factors that make one alternative preferred over another
might not be included in the model.
• If the solution is a unique optimal solution:
• It would be good to know how much worse the second-best solution is
than the unique optimal solution.
• If the second-best solution is very close to optimal:
• It might be preferred over the true optimal solution because of factors
outside the model.
68
Figure 12 - A Second-best Solution to the Ohio
Trust Location Problem

69