You are on page 1of 30



Nonlinear and



Learning Objectives
After studying this chapter, you will be able to:

Recognize when to use nonlinear optimization models. Recognize a quadratic optimization model.
Develop and solve nonlinear optimization models for Identify non-smooth optimization models and when
different applications. to use Evolutionary Solver.
Interpret Solver reports for nonlinear optimization. Formulate and solve sequencing and scheduling
Use empirical data and line-fitting techniques in models using Solvers alldifferent constraint.
nonlinear optimization.

M17_EVAN7821_09_SE_SUPPA_online.indd 1 03/10/14 12:48 PM

2 Supplementary Chapter ANonlinear and Non-Smooth Optimization

In many situations, the relationship among variables in a model is not linear.

Whenever either the objective function or a constraint is not linear, the model be-
comes a nonlinear optimization problem, requiring different solution techniques. In
a nonlinear optimization model (nonlinear program), the objective function and/or
constraint functions are nonlinear functions of the decision variables; that is, terms
cannot be written as a constant multiplied by a variable. Some examples of non-
linear terms are 3x 2, 4>y, and 6xy. Building nonlinear optimization models requires
more creativity and analytical expertise than linear or integer models that we have

Modeling and Solving Nonlinear Optimization Problems

Nonlinear models are generally more difficult to develop than linear models simply
because the objective function and constraint functions may not have the f amiliar linear
structure. Thus, building nonlinear optimization models relies on fundamental modeling
principles that we introduced in Chapter 11. For example, we may use b usiness logic to
identify the appropriate functional relationships among the decision variables or empiri-
cal data and line-fitting techniques to characterize the nonlinearities. In addition, non-
linear optimization models are considerably more difficult to solve than either linear or
integer models. However, Solver provides solution procedures that can solve nonlinear
optimization problems quite effectively.

A Nonlinear Pricing Decision Model

We introduced a simple nonlinear optimization model for finding the optimal price to
maximize revenue in Example 1.11. In this example, a market research study collected
data that estimated the expected annual sales for different levels of pricing. Analysts deter-
mined that sales can be expressed by the following model:
sales = -2.9485 * price + 3240.9
Using the fact that revenue equals price * sales, we express total revenue as
total revenue = price * sales
= price * 1-2.9485 * price + 3240.92
= -2.9485 * price2 + 3240.9 * price
Note that the total revenue function contains both a linear and a squared term for price,
making the model nonlinear.

Example A.1 Solving the Pricing Decision Model

A spreadsheet model for this problem was developed in should select Standard LSGRG Nonlinear to solve most
Chapter 2; this is shown again in Figure A.1. To find the nonlinear models. The Solver model is shown in Figure A.2.
maximum revenue using Solver, we identify the objective The optimal solution is to set the price at about $549 and
function cell as B11 and the decision variable cell as B5. achieve a total revenue of $890,574.
The model has no constraints. In Premium Solver, you

M17_EVAN7821_09_SE_SUPPA_online.indd 2 03/10/14 12:48 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 3

Figure A.1

Optimal Solver Solution for

Pricing Decision Model

Figure A.2

Solver Model for Pricing


The next example shows a more-complex pricing decision model involving multiple
decision variables.

Example A.2 A Hotel Pricing Model

The Marquis Hotel is considering a major remodeling rooms sold by 1.5%. Similarly, a 1% increase in the price
effort and needs to determine the best rates and room will decrease the number of rooms sold by 1.5%. For any
sizes to maximize revenues. Currently, the hotel has pricing structure (in $), the projected number of rooms
450 rooms with the following history: of a given type sold (we allow continuous values for this
example) can be found using the formula
Daily Avg.
Room Type Rate No. Sold Revenue historical average number of rooms sold +

Standard $85 250 $21,250 [(elasticity) (new price current price)

(historical average number of rooms sold)]
Gold $98 100 $9,800 (A.1)
current price
Platinum $139 50 $6,950
The hotel owners want to keep the price of a stan-
Total Revenue $38,000
dard room between $70 and $90; a gold room between
$90 and $110; and a platinum room between $120 and
Each market segment has its own price/demand elasticity.
$149. Define S = price of a standard room, G = price
Estimates are as follows:
of a gold room, and P = price of a platinum room. Then,
Room Type Price Elasticity of Demand for standard rooms, the projected number of rooms sold
is 250 [1.5(S 85)(250)] , 85 = 625 4.41176S. The
Standard 1.5
objective is to set the room prices to maximize total rev-
Gold 2.0 enue. Total revenue will equal the price multiplied by the
Platinum 1.0 projected number of rooms sold, summed over all three
types of rooms. Therefore, total revenue will be
This means, for example, that a 1% decrease in the
price of a standard room will increase the number of

M17_EVAN7821_09_SE_SUPPA_online.indd 3 03/10/14 12:48 PM

4 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Total Revenue The complete model is

= S(625 4.41176S) + G(300 2.04082G) + P(100 0.35971P )
maximize 625S + 300G + 100P 4.41176S2
= 625S + 300G + 100P 4.41176S2 2.04082G2 0.35971P2 2 2
2.04082G 0.35971P
To keep prices within the stated ranges, we need constraints: 70 " S " 90
70 " S " 90 90 " G " 110
120 " P " 149
90 " G " 110
1,025 4.41176S 2.04082G 0.35971P " 450
120 " P " 149
Figure A.3 shows a spreadsheet model (Excel file H otel
Finally, although the rooms may be renovated, there are Pricing Model) for this example with the optimal s olution.
no plans to expand beyond the current 450-room capacity. The decision variables, the new prices to charge, are given
Thus, the projected number of total rooms sold cannot ex- in cells B15:B17. The projected numbers of rooms sold are
ceed 450: computed in cells E15:E17 using the preceding formula.
By multiplying the number of rooms sold by the new price
(625 4.41176S) + (300 2.04082G)
for each room type, the projected r evenue is calculated, as
+ (100 0.35971P) " 450
given in cells F15:F17. The total revenue in cell F18 repre-
or sents the objective function.
Note that it is easier to formulate this model more as
1,025 4.41176S 2.04082G 0.35971P " 450 a financial spreadsheet than to enter the analytical formu-
las as they were developed. The Solver model is shown
in Figure A.4. The optimal prices predict a demand for all
450 rooms with total revenue of $39,380.65.

Figure A.3

Spreadsheet for Hotel

Pricing Model

M17_EVAN7821_09_SE_SUPPA_online.indd 4 03/10/14 12:48 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 5

Figure A.4

Solver Model for Hotel

Pricing Example

Interpreting Solver Reports for Nonlinear Optimization Models

Solver provides Answer, Sensitivity, and Limits reports for nonlinear optimization models.
However, the Sensitivity report is quite different from that for linear models. We use the
hotel pricing example to discuss these differences.

Example A.3 Interpreting Solver Reports for the Hotel Pricing Model
The Answer report, shown in Figure A.5, provides a change in the solution. Lagrange Multipliers in the
the same basic information as for linear models. The Constraints section are similar to shadow prices for lin-
Constraints section provides the value for the left-hand ear models. However, for nonlinear models, the Lagrange
side of each constraint in the Cell Value column, its bind- multipliers give only an approximate rate of change in the
ing or nonbinding status, and the value of the slack. In objective function as the right-hand side of a binding con-
this example, we see that the limit of 450 rooms and the straint is increased by 1 unit. Thus, for this example, if the
lower bound on the price of a gold room are binding. This number of available rooms is increased by 1 to 451, the
suggests that we could increase revenue if we could either total revenue would increase by approximately $12.08.
increase the capacity of the hotel or lower the minimum (For linear models, as we have seen, shadow prices give
price for a gold room. the exact rate of change within the Allowable Increase and
In the Adjustable Cells section of the Sensitivity Allowable Decrease limits.) Thus, you should be some-
report (Figure A.6), the Reduced Gradient is analogous what cautious when interpreting these values and will need
to the Reduced Cost in linear models. For this problem, to re-solve the models to find the true e ffect of changes
however, the objective function coefficient of each price to constraints. For this example, the o ptimal revenue
depends on many parameters; therefore, the reduced gra- for a 451-room capacity is $39,392.52, an increase of
dient is more difficult to interpret in relation to the prob- $39,392.52 $39,380.65 = $11.87, which is close to,
lem data. Thus, we cannot necessarily conclude that a but not exactly, the amount predicted by the Lagrange
decrease in the price of a gold room of $42.69 will force multiplier value.

M17_EVAN7821_09_SE_SUPPA_online.indd 5 03/10/14 12:48 PM

6 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.5

Hotel Pricing
Example Solver
Answer Report

Figure A.6

Hotel Pricing Example Solver

Sensitivity Report

Locating Central Facilities

A common problem in designing service systems is to locate a facility in a central
location with respect to other facilities to minimize some measure of distance from the
central location to each of the other facilities. Distances can be measured in several ways.
Suppose that the X- and Y-coordinates of two locations are (X1 , Y1 ) and (X2 , Y2 ). One
measure is the straight-line distance between the points. This is computed as

straight@line distance between 1X1, Y12 and 1X2, Y22 = 21X1 - X222 + 1Y1 - Y222 (A.2)

Equation A.2 is often called Euclidian distance. A second measure of distance is called
rectilinear distance, and it is computed as

rectilinear distance between 1X1, Y12 and 1X2, Y22 =  X1 - X2  +  Y1 - Y2   (A.3)

This is often called the city block metric. Rectilinear distance moves along the coor-
dinate axes; Figure A.7 illustrates the difference. Distances are often weighted based
on the volume, or frequency, of trips between locations.

M17_EVAN7821_09_SE_SUPPA_online.indd 6 03/10/14 12:48 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 7

Figure A.7

Straight-Line versus
Rectilinear Distance

Example A.4 Finding the Best Location for a Medical Laboratory

A medical testing laboratory needs to collect blood XC and YC represent the coordinates of the laboratory. We
samples from several regional hospitals to perform diag- want to minimize the weighted distance between the labo-
nostic testing. Currently, the laboratory is in the middle ratory and all locations, where the weights are the number
of a small town, but several new hospitals have been of trips/month:
constructed in outlying areas. The lab wants to relo-
cate to reduce the travel distance required to pick up minimize 5 2( X1 XC)2 + ( Y1 YC)2
the samples. The X- and Y-coordinates for the hospital + 252( X2 XC ) 2 + ( Y2 YC )2
locations have been found on a grid. The following table + 202( X3 XC ) 2 + ( Y3 YC )2
shows these, along with the average number of trips per
+ 352( X4 XC ) 2 + ( Y4 YC )2
month that the lab must make to each location.
+ 152( X5 XC ) 2 + ( Y5 YC )2
Location X-Coordinate Y-Coordinate Trips/month Figure A.8 shows a spreadsheet model (Excel file
Laboratory Location Model); the Solver model minimizes
1 0 0 5
the total distance in cell C20 by changing the decision
2 20 80 25 variables in cells B23 and C23. Figure A.9 shows the
3 60 30 20 hospital locations and recommended laboratory location
4 100 100 35 on a coordinate system using a bubble chart. The sizes
5 70 110 15 of the bubbles represent the numbers of trips/month,
and the star is the location of the laboratory for the op-
The area is quite rural, so the distances between timal solution. Note that the best location is centered
locations can be measured as the crow fliesthat is, us- closer to the locations with the highest number of trips
ing the straight-line distance formula (A.2). Suppose that per month.

M17_EVAN7821_09_SE_SUPPA_online.indd 7 03/10/14 12:48 PM

8 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.8

Laboratory Location Model


Figure A.9

Optimal Solution to the

Laboratory Location Model

M17_EVAN7821_09_SE_SUPPA_online.indd 8 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 9

The Economic Order-Quantity Model

In managing the inventory of retail goods such as grocery items, printer ink, and other
commodities that have stable demand over time, managers must make a fundamental de-
cision about how much to order and when to place orders. For example, they can order
a large quantity at one time or smaller quantities more frequently. These decisions have
economic impacts. Inventory costs fall into four major categories:
Purchase costsunit costs per item to purchase from suppliers
Order preparation costscosts involve the time spent preparing and placing
orders, such as clerical, telephone, receiving, and inspection time
Inventory-holding costall expenses associated with carrying inventory, such
as rent on storage space, utilities, insurance, taxes, and the cost of capital
Shortage costsadditional costs for shipping, invoicing, and labor for back
orders or lost profit opportunities and possible future loss of revenues because
of lost sales

Analytics in Practice: L
 ocating Collection Facilities
for the Red Cross
Two regional centers of the American Red Cross Blood
Program had the option of changing their main
warehouse/garage configuration and wanted to exam-
ine the effects of such a change.1 Both centers were in-
terested in considering either a single central site or a
combination of a central site with a satellite staging area
having limited facilities for storage of consumable sup-
plies, vehicle repairs, and so on. The problem was cast as
Littleny |

minimizing travel distances between points on a plane,

where the points represent the staging area and the col-
lection sites. Although a city may have numerous collec-
tion points scattered throughout its area, it was decided
to treat all collection points in a city as if they occurred at
the center of the city. This was consistent with the way
the participating blood centers kept mileage data on The users recognized that site analysis on the
operations; that is, only rough estimates were used. basis of distance traveled is only one facet of a site
Highway distances were estimated by treating the selection; cost per square foot, availability of sites,
highways as straight lines, whereas city travel was highway access, location of processing plants, and so
estimated by rectilinear distances. However, due to on, would ultimately have to be weighed in the evalua-
the urban/rural mix of the regions in this application, it tion. Therefore, minimal travel distance was not the sole
was decided to present the results of the study along determining factor for a location but was simply an input
both metrics on the assumption that the true distance into the decision process. As a result of the study, one
would be somewhere in between. Further, to obtain a region was able to locate suitable facilities near the opti-
reasonable estimate of travel distance, each collection mal point and was able to justify the move in part by the
point was weighted by the number of times it was vis- results of the study. Over time, the new location saved
ited during a year. costs of fuel and disposable supplies.

1Based on Robert P. Cerveny, An Application of Warehouse Location Techniques to Bloodmobile

Operations, Interfaces, 10. 6 (December 1980).

M17_EVAN7821_09_SE_SUPPA_online.indd 9 03/10/14 12:49 PM

10 Supplementary Chapter ANonlinear and Non-Smooth Optimization

These costs are influenced by the amount ordered and the timing of the ordering decision.
For example, if many small orders are placed, then the ordering cost will be high, but little
inventory will be carried, reducing holding costs. On the other hand, if few large orders
are placed, then ordering costs will be low, but inventory-holding costs will be high. Simi-
larly, if orders are placed too early, excessive holding will result. If orders are placed too
late, the firm risks running out of stock and incurring shortages. Thus, decision makers
seek a minimum-cost balance among these costs. The economic order quantity is the
amount to order that minimizes the total cost of ordering and holding.
To develop an EOQ model, we need to make several important assumptions:
1. Only a single inventory item is considered.
2. The entire quantity arrives at one time.
3. The demand for the item is constant over time.
4. No shortages are allowed.
We define the following variables:
Q = order quantity
D = annual demand
C = unit cost of the item
C0 = cost per order placed
i = inventory carrying charge per unit
With an annual demand D and order quantity Q, the number of orders that must be
placed each year is D>Q. Therefore, the total annual ordering cost is

annual ordering cost = 1number of orders per year21cost per order2 = 1D>Q2C0 (A.4)

Inventory holding cost per unit is computed as a fixed percentage of the unit cost
of the item. Thus, if C is the unit cost of the item and i is the annual carrying charge rate
(expressed as a decimal fraction), the cost of holding 1 unit for 1 year is
holding cost per unit = iC (A.5)
The assumption of constant demand becomes important in determining an expression
for the average number of units in inventory. If Q units are ordered and then received
when the inventory level is depleted, the inventory will decrease to zero at a constant
rate until the next order arrives. Because the inventory level starts at Q and ends at zero,
the average inventory is
Q+0 Q
average inventory = =  (A.6)
2 2
Thus, the annual inventory holding cost is

annual inventory holding cost = 1average inventory21annual holding cost per unit2

= a b iC =  (A.7)
2 2
We may now express the total annual inventory cost as
total annual cost = +  (A.8)
Q 2
Because Q is in the denominator of the order cost term, the model is nonlinear.

M17_EVAN7821_09_SE_SUPPA_online.indd 10 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 11

Example A.5 Solving the EOQ Model

To illustrate the EOQ model, suppose that annual A spreadsheet model (Excel file Economic Order Quantity
demand is 15,000 units, ordering cost is $200 per order, Model) with the optimal solution is shown in Figure A.10.
unit cost is $22, and the carrying charge rate is 20%. The basic model is given to the left; to the right, we set
Using the data we have for this specific example, we have up a table to chart the cost functions. We see that the
(15,000)(200) (0.2)(22)Q order cost decreases and the holding cost increases as
total annual cost = + the order quantity goes up. The Solver model is shown
Q 2
3,000,000 in Figure A.11. The solution is to order in q uantities of
= + 2.2Q about 1,168 units to minimize the total cost.

Figure A.10

of the EOQ

M17_EVAN7821_09_SE_SUPPA_online.indd 11 03/10/14 12:49 PM

12 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.11

Solver Model for

the EOQ Decision

One issue that the model does not address directly is the decision of when to order. The
decision to reorder is based on the level of inventory and is called the reorder point. If we
order too early, we will have more stock than we need and incur unnecessary holding costs.
If we order too late, we run the risk of running out of stock before the order is received. Ide-
ally, we would like the order to be received at the same time we run out of stock. Therefore,
the reorder point depends on how long it takes an order to be received after it is placed;
this is called the lead time. We should set the reorder point to be the inventory level that
provides enough stock to satisfy all demand during the lead time. If the demand rate, D, is
constant, as assumed, and the lead time is t, then the reorder point should be equal to Dt.
In the previous example, we have D = 15,000 units per year. Suppose that the lead time
is 1 week, or 0.0192 years. The demand during the lead time is 115,000210.01922 = 288
units. If we place an order when the inventory level reaches 288, then the order will a rrive
when the stock level falls to zero. Understanding the lead time and reorder point will
be important when we discuss how to incorporate uncertain demand into this model in
Supplementary Chapter B.

Using Empirical Data for Nonlinear Optimization Modeling

For many applications of nonlinear optimization, the form of the objective or constraint
functions are derived from empirical data. The next example shows how we might use
empirical data and line-fitting techniques in nonlinear optimization.

Example A.6 A Model for Advertising Strategy

The DTP Corporation has two major products. Marketing where X1 and X2 are the amount spend advertising prod-
analysts have conducted experiments to gather data on ucts 1 and 2, respectively. Suppose that the total adver-
the effect of media advertising on profits. These data are tising budget is set to $500,000 and that at least $50,000
shown in Figure A.12. Using the Add Trendline feature in must be spent on each product. How should the com-
Excel, we find that logarithmic functions can adequately pany allocate its advertising budget between the two
fit the data: products?
product 1 profit = 49.699 + 1.1568 ln( X1) This can be determined by formulating and solving
a nonlinear optimization model. The objective function
product 2 profit = 19.913 + 0.4177 ln( X2)
is the total profit as a function of the amount spent on

M17_EVAN7821_09_SE_SUPPA_online.indd 12 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 13

advertising each product, and the constraints reflect Figure A.13 shows a spreadsheet model for this prob-
the budget limitation and advertising requirements. The lem (Excel file DTP Corporation) with the optimal Solver
model is solution; the Solver model is provided in Figure A.14. The
solution specifies that $370,000 of the advertising budget
maximize total profit = 69.612 + 1.1568 ln(X1)
be allocated to product 1 and the remainder to product 2.
+ 0.4177 ln(X2)
X1 + X2 " $500,000
X1 # $50,000
X2 # $50,000

Practical Issues Using Solver for Nonlinear Optimization

Many nonlinear problems are notoriously difficult to solve. Solver cannot guarantee
finding the absolute best solution (called a global optimal solution) for all problems.
A local optimum solution is one for which all points close by are no better than the

Figure A.12

DTP Corporation Marketing


Figure A.13

Spreadsheet Model for DTP


M17_EVAN7821_09_SE_SUPPA_online.indd 13 03/10/14 12:49 PM

14 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.14

Solver Model for DTP


solution (think of the analogy of being at the top of a hill when the highest peak is on
another mountain). The solution found often depends a great deal on the starting solu-
tion in your spreadsheet. For complex problems, it is wise to run Solver from different
starting points. You should also look carefully at the Solver results dialog box when
the model has completed running. If it indicates Solver has found a solution. All con-
straints and optimality conditions are satisfied, then at least a local optimal solution has
been found. If you get the message Solver has converged to the current solution. All
constraints are satisfied. then you should run Solver again from the current solution to
try to find a better solution.

Quadratic Optimization

A special class of nonlinear optimization models is known as a quadratic programming

model. A quadratic optimization model is one that has a quadratic objective and all
linear constraints. Recall from algebra that a quadratic function is f 1x2 = ax 2 + bx + c.
Inother words, a quadratic function has only constant, linear, and squared terms.
Quadratic optimization models can be solved using the Standard LP/Quadratic solving
method within Solver.

The Markowitz Portfolio Model

The Markowitz portfolio model is a classic quadratic optimization model in finance
that seeks to minimize the risk of a portfolio of stocks subject to a constraint on the
p ortfolios expected return.2 The decision variables are the percent of each stock to
allocate to the portfolio. (You might be familiar with the term asset allocation model
that many financial investment companies suggest to their clientsfor example, main-
tain 60% equities, 30%bonds, and 10% cash.) Recall from Chapter 4 that we can
measure risk by the standard deviation, or, equivalently, the variance. In the Markowitz
model, the objective function is to minimize the risk of the portfolio as measured by
its variance. Because stock prices are correlated with one another, the variance of the
portfolio must reflect not only variances of the stocks in the portfolio but also the
covariance between stocks.

2H.M. Markowitz, Portfolio Selection, Efficient Diversification of Investments (New York: John Wiley &
Sons, 1959).

M17_EVAN7821_09_SE_SUPPA_online.indd 14 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 15

Define x j to be the fraction of the portfolio to invest in stock j. The variance of a portfolio
is the weighted sum of the variances and covariances:

variance of portfolio = a s 2i x 2i + a a 2s ij x ix j
k k
i=1 i=1 j7i

s 2i = the sample variance in the return of stock i
s ij = the sample covariance between stocks i and j

Example A.7 An Example of the Markowitz Model

Suppose an investor is considering three stocks in which Second, the portfolio must have an expected return of at
to invest. The expected return for stock 1 is 10%; for least 10%. The return on a portfolio is simply the weighted
stock 2, 12%; and for stock 3, 7%; and she would like an sum of the returns of the stocks in the portfolio. This re-
expected return of at least 10%. Clearly one option is to sults in the constraint
invest everything in stock 1; however, this may not be a
10x1 + 12x2 + 7x3 # 10
good idea because the risk might be too high. Research
has found the variance-covariance matrix of the individ- Finally, we will assume that we cannot invest negative
ual stocks to be the following: amounts:

x1, x2, x3 # 0
Stock 1 Stock 2 Stock 3
Stock 1 0.025 0.015 0.002 The complete model is
Stock 2 0.030 0.005 minimize variance = 0.025x21 + 0.030x22 + 0.004x23
Stock 3 0.004 + 0.03x1 x2 0.004x1 x3
+ 0.010x2 x3
Using these data and formula (A.9), the objective func-
x1 + x2 + x3 = 1
tion is
10x1 + 12x2 + 7x3 # 10
minimize variance = 0.025x12 + 0.030x22 + 0.004x23
+ 2(0.015)x1x2 + 2( 0.002)x1x3 x1, x2, x3 # 0
+ 2(0.005)x2x3

The constraints must first ensure that we invest 100%

of our budget. Because the variables are defined as frac-
tions, we must have

x1 + x2 + x3 = 1

Figure A.15 shows a spreadsheet model for this example (Excel file Markowitz
Model); Figure A.16 shows the Solver model. The minimum variance of the optimal port-
folio is 0.012.
The Solver Sensitivity report is shown in Figure A.17. As we noted, for nonlinear
models, the Lagrange multipliers are only approximate indicators of shadow prices. For
this solution, the Lagrange multiplier predicts that the minimum variance will increase by
63.2% if the target return is increased from 10% to 11%. If you re-solve the model, you
will find that the minimum variance increases to 0.020, a 66.67% increase.
Optimization models can and should provide decision makers with valuable insight
beyond simply finding an optimal solution. Using spreadsheet models and Solver, it is

M17_EVAN7821_09_SE_SUPPA_online.indd 15 03/10/14 12:49 PM

16 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.15

Markowitz Portfolio Model

Spreadsheet Implementation

Figure A.16

Solver Model for Markowitz

Portfolio Model

M17_EVAN7821_09_SE_SUPPA_online.indd 16 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 17

Figure A.17

Markowitz Model Sensitivity


easy to systematically vary a parameter of a model and investigate its impact on the solu-
tion. For example, we might be interested in understanding the relationship between the
minimum risk and the target return.

Example A.8 Analyzing Risk versus Reward

By changing the target return in Example A.7 and subject to a risk (variance) no greater than 1%. This form of
re-solving the model (or by using Solvers Parameter the model would be
Analysis feature), we can obtain the results shown in
maximize 10x1 + 12x2 + 7x3
Figure A.18. This clearly shows that the minimum variance
begins to increase at a faster rate as the target return x1 + x2 + x3 = 1
increases, indicating that the investor faces an increasing
risk in seeking higher returns. The investor faces a trade- 0.025x21 + 0.030x22 + 0.004x23 + 0.03x1x2
off between risk and reward; we address this issue further 0.004x1x3 + 0.010x2x3 " 0.01
in Chapter 16. x1, x2, x3 # 0
An alternative modeling approach would be to maxi-
mize the return subject to a constraint on risk. For example, In this case, we would have a linear objective function and
suppose the investor wants to maximize expected return a mixture of linear and nonlinear constraints.

Figure A.18

Risk-Return Analysis
for Markowitz Portfolio

M17_EVAN7821_09_SE_SUPPA_online.indd 17 03/10/14 12:49 PM

18 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Analytics in Practice: A
 pplying Nonlinear Optimization at Prudential Securities
Prudential Securities Inc. (PSI) created a mortgage- use a scenario analysis to generate an optimal portfo-
backed securities (MBS) department to develop lio composed of securities with widely differing char-
models for managing complex investments.3 The de- acteristics that matches the investors performance
partment developed a variety of analytical models, profile under a variety of interest rate scenarios and
including linear, integer, and nonlinear optimization employ a weighting scheme on the scenarios to reflect
models, to help value, trade, and hedge MBSs in in- the portfolio managers actual view of the direction of
ventory and c onstruct portfolios. In one example, future interest rates. The portfolio manager defines
nonlinear optimization models are used to construct the desired portfolio performance for each scenario.
optimal portfolios for clients, matching the clients in- An optimization technique bundles the different sce-
vestment performance profile with constraints u nder narios and leads to optimal structured portfolios that
a variety of interest rate scenarios. The model inputs meet specified performance targets while taking into
required from the client include the portfolio perfor- account possible interest rate movement. These mod-
mance target, the securities to consider, diversification els have been used hundreds of times per day by PSI
restrictions, and view of future interest rates. Analysts personnel.

Evolutionary Solver for Non-Smooth Optimization

As we noted in Chapter 13, such Excel functions as IF, ABS, MIN, and MAX lead to
non-smooth models. In the fixed-cost problem for K&L Designs in Chapter 15, we needed
binary variables to incorporate the fixed costs into the objective function and model the
logical conditions that ensure that the fixed cost will be incurred only if the production in
a period is positive. As we noted, we did this to preserve linearity of the constraints so that
we could use Solvers linear optimization method to find an optimal solution. H owever, it
might have been simpler to use an IF function in the spreadsheet to model the fixed costs.
Doing so, however, results in a non-smooth model that violates the linearity conditions
required for the linear optimization solution method used by Solver. Nevertheless, this can
simplify the modeling task, especially for nonanalytics professionals.
Problems that are nonsmooth or involve both nonlinear functions and integer
v ariables are usually difficult to solve using conventional techniques. To overcome
these limitations, new approaches called metaheuristics have been developed by re-
searchers. These approaches have some exotic names, including genetic algorithms,
neural n etworks, and tabu search. Such approaches use heuristicsintelligent rules for
systematically searching among solutionsthat remember the best solutions they find
and then modify or combine them in attempting to find better solutions. Solvers Stand-
ard Evolutionary algorithm uses such an approach. The evolutionary algorithm is avail-
able in both the standard Excel-supplied Solver as well as in the premium version.

Spreadsheet Models with Non-Smooth Excel Functions

In the K&L Designs example with fixed costs in Chapter 15, we could express the objec-
tive function as

minimize 11PA + 14PW + 12.50PS + 1.20IA + 1.20IW + 1.20IS + IF1PA 7 0,65,02

+ IF1PW 7 0,65,02 + IF1PS 7 0,65,02

3Based on Yosi Ben-Dov, Lakhbir Hayre, and Vincent Pica, Mortgage Valuation Models at Prudential
Securities, Interfaces, 22, 1 (JanuaryFebruary 1992): 5571.

M17_EVAN7821_09_SE_SUPPA_online.indd 18 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 19

Now, the only constraints needed are the material balance constraints:
PA - IA = 150
PW + IA - IW = 400
PS + IW - IS = 50
In this way, there is no need for the binary variables and the additional constraints that
involve them, which are more difficult to logically understand and model.
We illustrate Evolutionary Solver first using the new model for K&L Designs and
then with some other applications.

Example A.9 Using Evolutionary Solver for the K&L Design Fixed-Cost Problem
Figure A.19 shows a simpler, modified spreadsheet for the bounds to restrict the search space to a m anageable
K&L Designs fixed-cost problem (Excel file K&L Designs region. Thus, we set upper bounds of 600 (the to-
Evolutionary Solver Model). The objective function in cell B20 tal demand) and lower bounds of 0 for each of them.
is = SUMPRODUCT(B6:D9,B14:D15) + IF(B14+ 0,B9,0) + Evolutionary Solver finds essentially the same optimal
IF(C14 + 0,C9,0) + IF(D14+ 0,D9,0). Figure A.20 shows solution as the integer optimization problem we solved in
the Solver model. The Evolutionary Solver algorithm Chapter 15, except for a minor rounding issue in cell C14.
requires that all variables have simple upper and lower

Some applications use absolute value functions in the objective function or con-
straints. Absolute value functions are similar to IF functions and result in non-smooth
functions, necessitating the use of Evolutionary Solver.

Figure A.19

Modified Spreadsheet for

K&L Designs

M17_EVAN7821_09_SE_SUPPA_online.indd 19 03/10/14 12:49 PM

20 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.20

Evolutionary Solver Model

for K&L Designs

Example A.10 A Rectilinear Location Model

Edwards Manufacturing is studying where to locate a tool distance measure in formula (A.3) to compute the distance
bin on the factory floor. The locations of five production cells between locations. The optimal location should minimize
are expressed as X- and Y-coordinates on a rectangular grid the total weighted distance between the tool bin and all
of the factory layout. The daily demand for tools (measured production cells, where the weights are the daily number
as the number of trips to the tool bin) at each production cell of trips to the tool bin.
is also known. The relevant data are as follows: To formulate an optimization model for the best loca-
tion, define (X, Y) as the location coordinates of the tool
Cell X-Coordinate Y-Coordinate Demand bin. The weighted distance between the tool bin and each
cell is expressed by the objective function
Fabrication 1 4 12
Paint 1 2 24 minimize 12( X 1 + Y 4 ) + 24( X 1 + Y 2 )
Subassembly 1 2.5 2 13 + 13( X 2.5 + Y 2 ) + 7( X 3 + Y 5)
Subassembly 2 3 5 7 + 17( X 4 + Y 4 )
Assembly 4 4 17
The absolute value functions used in this objective func-
tion create a non-smooth model. Thus, Evolutionary Solver
Because of the nature of the equipment layout in the fac- is an appropriate solution technique.
tory and for safety reasons, workers must travel along Figure A.21 shows a spreadsheet model with
marked horizontal and vertical aisles to access the tool the o ptimal Evolutionary Solver solution for the
bin. Thus, the distance from a cell to the tool bin cannot E dwards Manufacturing example (Excel file Edwards
be measured as a straight line; rather it must be measured M anufacturing). The upper bounds are chosen as the
as rectilinear distancethat is, distances parallel to the maximum coordinate values and the lower bounds are
axes of the coordinate system. We may use the rectilinear zero. The Solver model is shown in Figure A.22.

M17_EVAN7821_09_SE_SUPPA_online.indd 20 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 21

Figure A.21

Edwards Manufacturing

Figure A.22

Evolutionary Solver Model

for Edwards Manufacturing

The results obtained by Evolutionary Solver can depend heavily on the starting values
of the decision variables and the amount of time devoted to the search. For simple problems,
it usually doesnt make much difference; however, for complex models, different starting
values can produce different results. In addition, increasing the maximum search time may
improve the solution. Thus, for complex problems, it is wise to run the p rocedure from
different starting points, and we suggest you try this on this example and end-of-chapter
problems. The maximum search time and other parameters can be changed from the
Options button in the dialog; however, this is usually only necessary for advanced users.

Optimization Models for Sequencing and Scheduling

A rather unique application of Excel modeling and Evolutionary Solver is for job-
sequencing problems. Job-sequencing problems involve finding an optimal sequence, or
order, by which to process a set of jobs. For any job sequence, we may compute the com-
pletion time for each job by successively adding the processing times of that job and all that

M17_EVAN7821_09_SE_SUPPA_online.indd 21 03/10/14 12:49 PM

22 Supplementary Chapter ANonlinear and Non-Smooth Optimization

were completed before it. We may then compare the completion times with the requested
due dates to determine if the job is either completed early or late. Lateness (L i) is the differ-
ence between the completion time 1Ci2 and the due date 1Di2, which can either be positive
or negative. Tardiness (Ti) is the amount of time by which the completion time exceeds the
due date; thus, tardiness is zero if a job is completed early). Hence, for job i,
L i = Ci - Di (A.10)
Ti = max50, L i6 (A.11)
Researchers have shown that sequencing jobs in order of shortest processing time
(SPT) first will minimize the average completion time for all jobs. Sequencing by ear-
liest due date (EDD) first will minimize the maximum number of tardy jobs. However,
the manager might be interested in minimizing other criteria, such as the average tar-
diness, total tardiness, or total lateness. It is possible to develop integer optimization
models for such problems, but these are quite complicated and can take a very long
time to solve.

Example A.11 Finding Optimal Job Sequences

Suppose that a custom manufacturing company has 10 processing time (Pi) and a due date (Di) that was requested
jobs waiting to be processed. Each job i has an estimated by the customer, as shown in the following table:

Job 1 2 3 4 5 6 7 8 9 10
Time 8 7 6 4 10 8 10 5 9 5
Due date 20 27 39 28 23 40 25 35 29 30

To develop a spreadsheet model for this problem, B13, =INDEX($B$4:$K$6,3,B10), finds the due date asso-
we use the Excel function INDEX to identify the pro- ciated with job 5.
cessing times and due date for the job assigned to a Any sequence of integers in the decision variable
particular s equence. Figure A.23 shows the model range is called a permutation. Our goal is to find a per-
(Excel file, Job Sequencing Model) and a portion of the mutation that optimizes the chosen criteria. Solver has
Excel formulas. A particular job sequence (the decision an option to define decision variables as a permutation;
variables) is given in row 10; for this example, we show this is called an alldifferent constraint. To do this, click
the sequence for the EDD rule. In rows 11 and 13, we on Integers in the Solver Parameters dialog, choose the
use the INDEX function to identify the processing time range of the decision variables, and then choose dif from
and due date associated with a specific job. For exam- the drop-down box as shown in Figure A.24. The final
ple, the formula in cell B11 is =INDEX($B$4:$K$6,2,B10). model, shown in Figure A.25, is quite simple: minimize the
This function references the value in the second row of chosen objective cellin this case, total tardinessand
the range B4:K6 corresponding to the job assigned to ensure that the decision variables are a valid permutation
cell B10, in this case, job 5. Likewise, the formula in cell of the job numbers. Figure A.26 shows the Solver solution.

The Traveling Salesperson Problem

The traveling salesperson problem (TSP) can be described as follows. A salesperson
needs to visit each of n different cities and return home in the minimum total distance. A
route that visits each city exactly once and returns to the start is called a tour. Many prac-
tical problems can be formulated as a TSP. For example, drivers for FedEx and UPS must
deliver packages to customers and return to their central location. Soft-drink vendors must

M17_EVAN7821_09_SE_SUPPA_online.indd 22 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 23

Figure A.23

Spreadsheet for Job

Sequencing Model

Figure A.24

Solver alldifferent Constraint


collect money and replenish bottles for a set of retail locations and then return to the ware-
house. Other examples are programming drilling machines to drill holes in circuit boards
and picking orders within a warehouse. In all these applications, the goal is to perform the
task in minimum total time or distance.

M17_EVAN7821_09_SE_SUPPA_online.indd 23 03/10/14 12:49 PM

24 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.25

Solver Model for Job

Sequencing to Minimize
Total Tardiness

Figure A.26

Evolutionary Solver
Solution for Minimum
Total Tardiness

In general, a TSP for n cities has 1n - 12! possible tours, making it quite difficult to
identify the optimal one. For example, if n is just 14, there are more than 6 billion possible
tours! The use of algorithms such as Evolutionary Solver allow us to find near-optimal
solutions efficiently.

Example A.12 Touring American Baseball League Cities

An avid baseball fan in Detroit would like to plan a trip to capabilities of Solver, as we soon explain). Any city can be
visit all 14 ballparks of teams in the American Baseball chosen as the start because a feasible sequence requires
League and minimize the total travel distance. Figure A.27 that the fan visit each city exactly once and must return
shows a spreadsheet model designed for Evolutionary to the starting point, so we arbitrarily choose city 0 as the
Solver (Excel file American Baseball League). The data start. The decision variables are in the range B23:B35; the
matrix shows the distances between each pair of cities. tour shown in the figure sequences the cities in numeri-
The model is somewhat tricky to develop and requires cal order. Note that when we get to the (n 1)st city, we
some detailed explanation to understand. We number must return to city 0, so cell B36 is not a decision variable.
the cities from 0 to 13 (this is necessary to use one of the We set cell A23 to 0. Whatever city is chosen to move to

M17_EVAN7821_09_SE_SUPPA_online.indd 24 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 25

from city 0 is the decision variable in cell B23. We must visited in the tour from city 0 exactly once. This is accom-
ensure that this city becomes the From city in the next plished by using the alldifferent constraint, as we saw in
row; thus, the formula in cell A24 is =B23. In other words, the job-sequencing example. Because the alldifferent con-
we simply copy the value of cell B23 into cell A24 and do straint applies to a set of positive integers from 1 to N, we
this for the remaining cells in column A. The INDEX func- needed to designate the first city as 0 (otherwise the deci-
tion is used to find the distance from one city to another sion variables would have had to range from 2 to 14, and we
(note that the INDEX function refers to the row and column would not have been able to use the alldifferent constraint).
numbers of the range $D$5:$Q$18 and not the numbering Figure A.29 shows the solution that Evolutionary
of the cities that we used). Finally, in columns E and F we Solver finds. To find the tour starting in Detroit, begin at
use the VLOOKUP function to translate the numerical city row 34 and cycle around the tour to Milwaukee, Minnesota,
values in the tour to their names. Seattle, Oakland, California, Texas, Kansas City, Chicago,
Figure A.28 shows the Solver model. The objective is Cleveland, Baltimore, New York, Boston, Toronto, and r eturn
to minimize the total tour length in cell C37 by changing the to Detroit, for a total distance of 6,718 miles. Of course, the
decision variables in the range B23:B35. The key to using real problem should incorporate the game schedules (who
the Standard Evolutionary algorithm is to ensure that the wants to visit a ball park if the team is out of town?), but this
decision variables include each of the remaining 13 cities would require a much more complicated model.

Figure A.27

Spreadsheet Model for

the American Baseball
League TSP

M17_EVAN7821_09_SE_SUPPA_online.indd 25 03/10/14 12:49 PM

26 Supplementary Chapter ANonlinear and Non-Smooth Optimization

Figure A.28

Solver Model for the

American Baseball
League TSP

Figure A.29

Solver Solution for the

American Baseball League

Key Terms

Economic order-quantity Permutation

Euclidian distance Quadratic optimization model
Global optimal solution Rectilinear distance
Lagrange multiplier Reduced gradient
Lead time Reorder point
Local optimal solution Tour
Nonlinear optimization model Traveling salesperson problem (TSP)
(nonlinear program)

M17_EVAN7821_09_SE_SUPPA_online.indd 26 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 27

Problems and Exercises

Note: Data for many of these problems are provided For the pricing decision model in Example A.1, sup-
in the Excel file Chapter A Problem Data to facilitate pose that the company wants to keep the price at a
model building. Worksheet tabs correspond to problem maximum of $500. Note that the solution in
scenarios. Figure A.1 will no longer be feasible. Modify the
spreadsheet model to include a constraint on the
Problem 15 in Chapter 1 posed the following situa-
maximum price and solve the model. Interpret the in-
tion: A manufacturer of mp3 players is preparing to
formation in the Solver Sensitivity report.
set the price on a new model. Demand is thought to
depend on the price and is represented by the model In the hotel pricing problem in Example A.2, sup-
pose that the hotel is considering adding suites to its
D = 2,500 - 3P
room mix. Based on an analysis of local competitors,
The accounting department estimates that the total
suites can sell for a rate of $180, and they expect to
costs can be represented by sell 20 per day to business travelers. The price elas-
ticity of demand is estimated to be 2.5. The hotel
C = 5,000 + 5D
would want to keep the price of suites between $150
You were asked to develop a model for the total
and $200. Modify the spreadsheet to include suites
profit and implement it on a spreadsheet. Use non- and find prices that will maximize total revenue.
linear optimization with Solver to find the price that
A franchise of a chain of Mexican restaurants wants
maximizes profit.
to determine the best location to attract customers
Problem 16 in Chapter 1 posed the following situa-
2. from three suburban neighborhoods. The coordinates
tion: The demand for airline travel is quite sensitive of the three suburban neighborhoods are as follows:
to price. Typically there is an inverse relationship
between demand and price; when price decreases, Neighborhood X-Coordinate Y-Coordinate
demand increases, and vice versa. One major airline Liberty 2 12
has found that when the price (p) for a round trip be-
Jefferson 9 6
tween Chicago and Los Angeles is $600, the demand (D)
is 500 passengers per day. When the price is reduced Adams 1 1
to $400, demand is 1,200 passengers per day. You
were asked to develop an appropriate model. Use The population of Adams is four times as large as

nonlinear optimization with Solver to find the opti- Jefferson, and Jefferson is twice as large as Liberty.
mal price to maximize revenue. The restaurant wants to consider the population in
its location decision. Develop and solve a model to
Problem 8 in Chapter 11 posed the following situation:
3. find the best location, assuming that straight-line dis-
The Radio Shop sells two popular models of portable tances can be used between the locations.
sport radios, model A and model B. The sales of these
products are not independent of each other (in eco- ElectroMart wants to identify a location for a ware-
nomics, we call these substitutable products, because house that will ship to five retail stores. The coor-
if the price of one increases, sales of the other will dinates and annual number of truckloads are given
increase). The store wishes to establish a pricing here. Develop and solve a model to find the best
policy to maximize revenue from these products. A location, assuming that straight-line distances can be
study of price and sales data shows the following re- used between the locations.
lationships between the quantity sold (N) and prices
(P) of each model: Retail Store X-Coordinate Y-Coordinate Truckloads
A 18 15 12
NA = 20 - 0.62PA + 0.30PB
B 3 4 18
NB = 29 + 0.10PA - 0.60PB C 20 5 24
You were asked to construct a model for total rev-
D 3 16 12
enue and implement it on a spreadsheet. Use non- E 10 20 18
linear optimization with Solver to find the optimal
prices to maximize revenue.

M17_EVAN7821_09_SE_SUPPA_online.indd 27 03/10/14 12:49 PM

28 Supplementary Chapter ANonlinear and Non-Smooth Optimization

The annual demand for the bolt-nut package for

8. 11. In Chapter 8 we noted that the least squares coef-
the aircraft manufacturer Excel database Purchase ficients in a regression model are found by mini-
Orders (see Example 1.3 in Chapter 1) is 60,000. mizing the sums of squares of the errors as given
The fixed cost of placing an order is $50 and the in equation (8.4). This is a nonlinear optimization
unit cost of the item is $3.95. The company uses an problem. Using the Home Market Value data, set
annual carrying charge of 15%. Use the Economic up a spreadsheet model to find the values for the
Order Quantity Model spreadsheet and Solver to find slope and intercept using nonlinear optimization.
the optimal order quantity. Compare your answer with Example 8.4.
The Economic Order Quantity (EOQ) model is based
9. 12. Many manufacturing situations, for example, the pro-
on ordering only a single item. Suppose we allow duction of such large and complex items as aircraft
for multiple items that are independent except for a or machines, exhibit a learning effect in which the
budget restriction. Develop a model to minimize the production time per unit decreases as more units are
total cost and meet a budget constraint on the total produced. This is often modeled by a power curve,
inventory value, defined by the product of the unit y = ax -b, where a and b are constants. Suppose that
cost multiplied by the order quantity, summed over data on production times for the first 10 units pro-
all items. Apply your model to the following data duced were collected from a new project at Glasgow
by implementing it on a spreadsheet and solving the Machine Tool:
model with Solver. Interpret the sensitivity report.
Unit Production Hours
Item 1 Item 2 Item 3 1 3,161
Annual demand 2,000 2,000 1,000 2 2,720
Item cost $100 $50 $80 3 2,615
Order cost $150 $135 $125 4 2,278
Budget limitation = $20,000 5 2,028
i = 0.20 6 2,193
7 2,249
10. A manufacturer stocks four items, each requiring a
8 2,268
certain amount of floor space in the warehouse. A
limited number of square feet of floor space is avail- 9 1,994
able to store the total quantity of the four products 10 2,000
that are ordered. Develop a model based on the eco-
nomic order quantity to minimize the total cost and Develop a model for estimating the power curve

meet the floor space limitation. Apply your model to to minimize the sum of the squared deviations of
the following data by implementing it on a spread- the errors. Use nonlinear optimization to find the
sheet and solving the model with Solver. Interpret the parameters.
Sensitivity report. Assume that the carrying charge
13. For the DTP Corporation data in Example A.6, fit a
rate is 20% and that 500 square feet of floor space
are available for storage. second-order polynomial function to model the profit
for each product. Use these functions to find the opti-
Item 1 Item 2 Item 3 Item 4 mal advertising allocation with the same constraints as
the example. Interpret the Solver Sensitivity report.
Annual demand 5,000 1,000 7,500 6,000
Ordering $75 $50 $50 $75 14. The Hal Chase Investment Planning Agency is in
cost/order business to help investors optimize their return from
investment. Hal deals with three investment medi-
Unit cost $100 $300 $500 $200
ums: a stock fund, a bond fund, and his own Sports
Floor space/unit 2 1.5 3 1 and Casino Investment Plan (SCIP). The stock fund is
(square feet) a mutual fund investing in openly traded stocks. The

M17_EVAN7821_09_SE_SUPPA_online.indd 28 03/10/14 12:49 PM

Supplementary Chapter ANonlinear and Non-Smooth Optimization 29

bond fund focuses on the bond market, which has formulation. Modify the spreadsheet and solve with
a more stable, but lower, expected return. SCIP is a Evolutionary Solver. Compare your solution with the
high-risk scheme, often resulting in heavy losses but optimal solution in Figure 15.23.
occasionally coming through with spectacular gains.
16. Formulate Chris Corrys investment problem (Prob-
Average returns, their variances, and covariances are
lem 23 in Chapter 15) on a spreadsheet without using
as follows:
linking constraints for the binary variables but using
IF statements instead. Modify the spreadsheet and
Stock Bond SCIP
solve with Evolutionary Solver.
Average return 0.148 0.060 0.152
17. Modify the economic order quantity problem in this
Variance 0.014697 0.000155 0.160791
chapter (Example A.5) to include discounts. Assume
Covariance with 0.000468 0.002222
that the unit cost is $22 for the first 1,000 units in
an order but drops to $19 for any additional units or-
Covariance with 0.000227 dered. Use IF statements in the spreadsheet model to
bond incorporate the quantity discount and Evolutionary
Solver to find the best solution.
Develop and solve a portfolio optimization model for

this situation for a target return of 12%. 18. Stout Investments wishes to design a minimum vari-
ance portfolio of index funds. The funds selected for
15. Modify the plant location model in Chapter 15 consideration and their variance-covariance matrix
(E xample 15.10) to use IF statements to express and average returns are given next:
the linking constraints in the mixed-integer model

Bond S&P 500 Small Cap Mid Cap Large Cap Market Commodity
Bond 0.002%
S&P 500 0.001% 0.020%
Small cap 0.001% 0.027% 0.047%
Mid cap 0.001% 0.024% 0.039% 0.033%
Large cap 0.001% 0.019% 0.027% 0.023% 0.027%
Emerging market 0.000% 0.032% 0.050% 0.043% 0.041% 0.085%
Commodity 0.000% 0.000% 0.005% 0.005% 0.009% 0.015% 0.054%
Average weekly return 0.044% 0.118% 0.256% 0.226% 0.242% 0.447% 0.053%

a. Formulate and solve a Markowitz portfolio opti- How would the optimal portfolio change? Compare

mization model for this situation having a target the solutions obtained using the nonlinear optimiza-
return of 0.19%. tion and Evolutionary Solver options.
b. Suppose the company wants to restrict the per- 19. An IT support group at Thomson State College has
centage of investments in each fund as follows: seven projects to complete. The time in days and
Bond: between 10% and 50% project deadlines are shown next.
S&P 500: between 30% and 50%
Small cap: no more than 20% Project 1 2 3 4 5 6 7

Mid cap: no more than 20% Time 4 9 12 16 9 15 8

Large cap: no more than 20% Deadline 12 24 60 28 24 36 48

Emerging market: no more than 10%

Commodity: no more than 20%

M17_EVAN7821_09_SE_SUPPA_online.indd 29 03/10/14 12:49 PM

30 Supplementary Chapter ANonlinear and Non-Smooth Optimization

a. Sequence the projects to minimize the average 20. Suppose the distances that a pharmaceutical rep-
lateness. resentative, Tracy Ross, travels between medical
b. Sequence the projects to minimize the average offices are
c. Compare these solutions to the SPT and EDD

From 1 2 3 4 5 6 7 8
1 0 19 57 51 49 4 12 92
2 19 0 51 10 53 25 80 53
3 57 51 0 49 18 30 6 47
4 51 10 49 0 50 11 91 38
5 49 53 18 50 0 68 62 9
6 4 25 30 11 68 0 48 94
7 12 80 6 91 62 48 0 9
8 92 53 47 38 9 94 9 0

Set up and solve a traveling salesperson problem using Evolutionary Solver.

Case: Performance Lawn Equipment

The accounting department at PLE reviewed the that the following nonlinear function could model the rate-
production-planning model assumptions in the case in change costs rather accurately:
Chapter 14 and does not believe that the rate-change cost
rate@change cost = 0.0051Xt - Xt-122
was appropriately modeled using a fixed $5.00 per unit.
Instead, their analysis shows that the rate-change cost Modify your model and analysis from the case in
grows faster as the deviation from production in the pre- Chapter 14 using this assumption. Summarize your results
vious month increases. Specifically, their analyst found in a short memo to Ms. Burke.

M17_EVAN7821_09_SE_SUPPA_online.indd 30 03/10/14 12:49 PM