ss#4 Sol

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Solutions to Self Study Problem Set #4

1. Bidding Simulation
There are two sources of uncertainty in this problem: the bid of Incredible Edibles and the bid of Hot
Potato. Let I be a random variable representing the bid of Incredible Edibles. Similarly, let H be a
random variable representing the bid of Hot Potato. The problem dictates that I is unformly distributed
between $11,000 and $13,000 (10% to 30% above the $10,000 cost), and H is unformly distributed
between $11,500 and $12,500 (15% to 25% above the $10,000 cost). Equivalently, we could express
this using the following compact notation:
I ~ U[$11,000 , $13,000] and H ~ U[$11,500, $12,500].
Tasty Delight only wins the contract if its bid is the lowest of the three companies. Let T represent the
bidding policy, i.e., the number of dollars the company decides to bid, and let P be Tasty Delights
profit. Then we can write:
if T < min{I,H} then P = T - $10,000,
otherwise P = $0.

Decision Models & Optimization

Page 1

Term II, 2016

Solutions to Self Study Problem Set #4


2. Comparing One Fast Machine with 10 Slow Ones
The spreadsheet acmeCB.xls contains a sample implementation of this problem.
a) Model (one sample of) the production cost obtained when operating the new facility.
Let X equal the potential production in one day of operation of the new facility. We define X to be a
normally distributed random variable with mean 35 and standard deviation 7 (X ~ N(35,7)). Then we
compute daily cost as follows:
step 1: let x = X (i.e., sample X and store the value in x)
Since we can shut the plant down to avoid overproduction . . .
step 2: let actual production, A = min{x,35}
step 3: let daily cost = $10 * (35 A)
b) Model (one sample of) the production cost obtained when operating the 10 old lines.
We use the same methodology as in (a). Let X1, X2, ... , X10 be the potential daily production on
each of 10 lines. We define each Xi to be normally distributed with mean 3.5 and standard deviation
1.05 (i.e., Xi ~ N(3.5,1.05)), and note that the sampled value of each lines potential production
should be independent of that of the others. Here we compute the daily cost in 4 steps:
step 1: sample each Xi ; that is, let xi = Xi for i = 1,...,10
step 2: let daily potential production equal P = x1+ x2+ ... + x10
step 3: let actual daily production A = min{35,P}
step 4: let daily cost = $10 * (35 A)
Again, we do not include the cost of producing too much, since we assume that as soon as
production for a day hits 35 tons, we can shut down all of the lines.
c) How much would Acme be willing to pay for the new facility?
Acme would not be willing to pay anything for the new facility. The expected cost of running the old
lines ($12.26) appears to be less than that of running the new plant ($27.30).
d) Can you deduce from the simulation results how much Acme would be willing to pay each day to
lease a storage facility that would allow it to store the excess inventory solvent that it produced? If
so, how much? If not, what can you say about the value?
The answer is unclear from the data given. Since we assumed the plant shut down as soon as the
required 35 tons were produced, we have no data on how much extra inventory tends to be
produced or, once produced, how quickly it gets used. We would need to create a new simulation
model to evaluate this new situation. What we do know is that the expected cost without an
inventory facility equals $12.26, so $12.26 is the most Acme would be willing to pay each day for a
facility under any circumstance.

Decision Models & Optimization

Page 2

Term II, 2016

Solutions to Self Study Problem Set #4


3. RollerDice
Excel File: RollerDice.xls
(a) Suppose the demand on a given evening turns out to be 120 regular and 35 luxury. What is the total revenue
that RollerDice will receive? Show all intermediate calculations.

revenue

= $120 min{140,120} + $360 min{30,35}


= $120 120 + $360 30
= $25,200

(b) Set up a simulation model to evaluate the revenue of this policy. That is, specify the random variables to be
sampled, their distribution types, their parameters, and correlations. Specify also formula(e) that should be
used to calculate the revenue.

Random Variables (assumption cells in Crystal Ball):


DR

= demand for regular rooms


~ normally distributed: N( = 100, = 35)

DL

= demand for luxury rooms


~ normally distributed: N( =30, = 8)

correlation: (DR , DL) = 0.3

Decision Variables
QR

= number of regular rooms to build

QL

= number of luxury rooms to build

Payout Calculation (forecast cell in Crystal Ball):


Revenue

= $120 min{QR , DR} + $360 min{QL , DL}

(Remark: Since the actual demand values should be integral and non-negative, we could easily
incorporate these restrictions into our model.
For example, define the actual demand for regular rooms as:

DR = max{round(DR), 0}.
However, the results are relatively insensitive to these modifications.)
Decision Models & Optimization

Page 3

Term II, 2016

Solutions to Self Study Problem Set #4


(c) Run the simulation for 500 trials, using a random number seed of 123 and choose Monte Carlo for the
sampling method. Print out a copy of the summary statistical data and frequency chart.

Frequency Histogram of 500 Trials


16%

percent of outcomes

14%
12%
10%
8%
6%
4%
2%

>2

7,

00

00

-2

25

,5
0

-2

5,

7,

50

00

,0
0
24

22

,5
0

-2

2,

4,

50

0
00
1,
,0
0

,5
0

21

19

-2
0

-2
0

-1

95

00

00

80

-1

,0
0

,5
0

00

16

-1

-1
0

5,

,1

,5
0

18

5,

6,
50

00

50

3,

00

13

,0
0

-1

2,

0
12

,5
0

50

10

9,

-1

0,
50

,0

-1

-9
0

7,

50

<=
7

,5

00

00

0%

$ Revenue

(d) Based on the results of the 500 trials, what is your estimate of the probability that revenue will exceed
$27,000?

From the frequency histogram, we estimate that the probability is about 11%.
Decision Models & Optimization

Page 4

Term II, 2016

Solutions to Self Study Problem Set #4


(e) Construct a 90% confidence interval for the expected revenue. Show all intermediate calculations. (The
appropriate z-value for a 2-sided, 90% confidence interval is 1.645.)

Confidence interval:
[sample average 1.645 mean standard error]
= $21,390.24 1.645 $202.76
[$21,056 , $21,724]

(f) Suppose that RollerDice wants a 90% confidence interval for the expected revenue with a width of $30. How
many simulation trials would be needed to give a 90% confidence interval of the form [a, b], where b - a = 30?
Briefly explain your answer.

Confidence interval = 30
Therefore, the width of [sample average 1.645 (4,533.91/n1/2)] must be 30.
Therefore, 1.645 (4,533.91/n1/2) = 15.
Solving for n,

n 247,000.
(g) RollerDice would like to know how likely it is to lose one or more customers of either type (regular or luxury)
because the desired room is not available. Indicate how to modify the model of part (b) to estimate this
probability. Re-run the simulation with the same parameters and report this probability. Give a 90%
confidence interval for this probability. Show all intermediate calculations.

Define:

L = indicates if any customers were lost


Include this additional calculation:
if {DR > QR or DL > QL} then L = 1 otherwise L = 0
Then, in addition to revenue, track L in each trial of the simulation (make it a forecast cell in Crystal
Ball). The average value of L over all of the simulation trials will equal the fraction of trials in which
customers were lost. This is the basis of our estimate of the probability that customers will be lost.
We have re-run the simulation to find an estimate of this probability. Below we show the summary
statistics from which you may estimate the probability.

Decision Models & Optimization

Page 5

Term II, 2016

Solutions to Self Study Problem Set #4

Note that our estimate of the probability that customers will be turned away, 0.49, is an estimate of
a mean. Furthermore, we know that the true mean is roughly normally distributed with a mean of
0.49 (our estimate of the mean) and a standard deviation of 0.0224 (the standard error of the
estimate of the mean). Using a z-value of 1.645 (see part (e) of this question) we can calculate the
90% confidence interval for the true probability.
Confidence interval:
[sample average 1.645 mean standard error]
= 0.4900 1.645 0.0224
[0.4532, 0.5269]
(h) Because of the design of the floors, the number of regular rooms must be a multiple of ten (and the number of
luxury rooms a multiple of five). For example, the hotel could be built with 100 regular rooms and 50 luxury
rooms, or 110 regular rooms and 45 luxury rooms, etc. Suppose that RollerDice wants to find the number of
regular and luxury rooms to build to maximize the expected revenue from the 60,000 square feet available.
List
(1) all of the possible room allocations RollerDice might consider;
(2) all of the reasonable room allocations that need to be tested.
Be specific in your reasons for excluding the room allocations not listed in (2).

Recall that the hotel has 60,000 square feet for rooms, that a regular room requires 300 square
feet, and that a luxury room requires 600 square feet.
(1) The maximum number of regular rooms is 200 (=60,000/300). Reducing regular rooms in
increments of 10 and increasing luxury rooms in increments of 5, we enumerate all 21 feasible
ways to build the hotel:
Numbers of Rooms
Regular

200

190

180

170

...

30

20

10

Luxury

10

15

...

85

90

95

100

Decision Models & Optimization

Page 6

Term II, 2016

Solutions to Self Study Problem Set #4


(2) Lets see if we can exclude some of the more extreme allocations of rooms.
Demand for both types of rooms is normally distributed, and any configuration of rooms will
have either more luxury rooms than the mean demand for luxury rooms, or more regular rooms
than the mean demand for regular rooms, or both. It would never make sense to err by building
fewer than the mean number of luxury rooms demanded, because the revenue contribution of a
luxury room per square foot is 50% higher than that of a regular room. We should build no
fewer than 30 luxury rooms.
How many more than 30? Should we allocate the entire space to luxury rooms? No, since the
regular rooms are quite profitable and the demand for luxury rooms will be above 45 rooms only
about one night per month. It is clearly unwise to build more than 45 luxury rooms.
We might therefore reasonably restrict ourselves to the following four cases:
Numbers of Rooms

Decision Models & Optimization

Regular

140

130

120

110

Luxury

30

35

40

45

Page 7

Term II, 2016

Solutions to Self Study Problem Set #4


4. Betting Simulation Problem
The spreadsheet bettingCB.xls contains a sample implementation of the problem.

a) Assume that the casino doesn't mind large, small, or odd-sized bets and doesn't notice the error in
the wheel until 400 spins are over. Model the payoff at the end of the day at the Wheel for Abbott
and for Costello. Assume they each start with $100.
Let Bi, i = 1, ..., 400 be 400 random variables that represent the outcomes of the 400 bets of one
day. The Bis are i.i.d., with the following distribution:

1 with probability 0.52 (win )


Bi
1 with probability 0.48 (lose )
Note that Abbott and Costello are both betting on White, so at each spin of the Wheel they both win
or lose. (The only difference between their betting strategies is the fraction of wealth that is bet.)
Therefore for each spin of the Wheel we use the same sampled value of Bi to calculate whether
both of them win or lose. Let bi be the sampled value of Bi.
Let Ai and, Ci (respectively) equal Abbott's and Costello's wealth after bet i, where A0 = C0 = $100 is
their initial wealth. Then we have
round 1)

b1 = B1
A1 = A0 * (1 + 0.03 * b1)
C1 = C0 * (1 + 0.20 * b1)

round 2)

b2 = B2
A2 = A1 * (1 + 0.03 * b2)
C2 = C1 * (1 + 0.20 * b2)

round 400)

b400 = B400
A400 = A399 * (1 + 0.03 * b400)
C400 = C399 * (1 + 0.20 * b400)

b) Using the results from the 500-trial simulation, construct a 90% confidence interval for the expected
final wealth under each strategy.
The 90% confidence interval for the mean can be calculated as X 1645
.
/ n , where X , , and
n represent the mean, standard deviation, and sample size of the simulation results. Then from the
statistical results provided with the question we can calculate:

Upper Limit of Confidence Interval


Lower Limit of Confidence Interval

Abbotts
Wealth
$ 170.92
$ 155.46

Costellos
Wealth
$730.58
$ 83.34

The question of whether to pick Abbott's or Costello's betting strategy depends, in the short run,
primarily on how risk-averse you are. Choosing Abbott's conservative strategy produces smaller
gains, but does so more consistently and with much less variance. Costello's bold strategy
occasionally produces incredibly large gains, but most often results in losing almost all his money.
In fact, the frequency table shows that Costello's final wealth is under $18.69 80% of the time.

Decision Models & Optimization

Page 8

Term II, 2016

Solutions to Self Study Problem Set #4


Some Interesting Facts about the Betting Problem (for the curious)
Researchers in the area of investment under uncertainty use a rule called the Kelly Criterion (after John
Kelly), that describes a rule for how much of your wealth you ought to bet in a situation in which you
have a given percentage advantage. The Kelly Criterion assumes that our objective function is the
growth rate of our money; the Kelly betting strategy thus prescribes the optimal fraction of your bankroll
to bet in each situation, to maximize the expected growth rate of our money.
The heuristic form of the Kelly Criterion is Edge over Odds. That is, bet a fraction of your bankroll
equal to the advantage you have when compared to a fair game (here, 2%-our 52% chance of winning
minus the 50% dictated by the 1-to-1 payoff) divided by the odds of winning (here, 52%). Thus, the bet
fraction that maximizes the expected growth rate of the initial $100 is 0.02/0.52, or 3.85% clearly much
closer to Abbott's strategy than Costellos. Over a very long period of time, money wagered at this 2%
advantage in 3.85% chunks will grow faster than money wagered in any other size chunks, either
larger or smaller. [In finance circles this is known as the OGGP, or Optimal Geometric Growth
Portfolio.] Note that the Kelly Fraction for negative-expectation gambles (such as craps, roulette, and
baccarat) is zero.
For more on the Kelly criterion, you can read Edward O. Thorps The Mathematics of Gambling (Lyle
Stuart; 1984) or Richard Epsteins Theory of Gambling and Statistical Logic (Academic Press, 1995).

Decision Models & Optimization

Page 9

Term II, 2016

Solutions to Self Study Problem Set #4


5. Bond Return Simulation (OPTIONAL)
The spreadsheet yieldCB.xls contains a sample implementation of this problem.
a) Model (one sample of) the returns for the bonds at the end of the year.
We use the formula given in the homework to calculate the 1-year return. Let N be the maturity of
the bond. Then with the current 8% yield, we can calculate today's price for all of the bonds: P0 =
100/1.08N.
Similarly, we can use the formula to price all of the bonds one year from now. The problem states
that all bonds will have the same yield next year, but that next year's yield is uncertain (it is normally
distributed with mean 8% and standard deviation 1%). Let Y ~ N(0.08,0.01) equal the yield next
year. For any trial of the simulation we sample Y once, setting y = Y. We then use the same
sampled value, y, to compute the price of all of the bonds in one year. In particular, for an N-year
bond in one year there will be N - 1 years to maturity. Then letting P1 equal the price of the bond in
one year, we have P1 = $100/(l+y)N-1.
Finally having computed P0 and P1 for each of the bonds, we can calculate each bond's return as
((P1-P0)/P0)*100%. To summarize:
Step 1:

P0,2-year = $100/1.082
P0,5-year = $100/1.085
P0,10-year = $100/1.0810
P0,20-year = $100/1.0820
P0,30-year = $100/1.0830

Step 2: let y = Y (sample Y); then

P1,2-year = $100/(1+y)1
P1,5-year = $100/(1+y)4
P1,10-year = $100/(1+y)9
P1,20-year = $100/(1+y)19
P1,30-year = $100/(1+y)29
Step 3: 2-year return = ((P1,2-year - P0,2-year)/ P0,2-year) * 100%
5-year return = ((P1,5-year - P0,5-year)/ P0,5-year) * 100%
10-year return = ((P1,10-year- P0,10-year)/ P0,10-year) * 100%
20-year return = ((P1,20-year - P0,20-year)/ P0,20-year) * 100%
30-year return = ((P1,30-year- P0,30-year)/ P0,30-year) * 100%
b) Comment on the following statement: "The expected yield of the 30-year bond one year from today
is 8%. At that yield, its price would be $10.73. The return for the year would be 8% =
(10.73-9.94)/9.94. Hence, the average return for the bond should be 8% as wella simulation isn't
really necessary."
This statement is false. As you see from the simulation, the 95% confidence intervals for the return
on the 30-year bond do not include the value 8.0%. This is a strong indication that the expected
returns are indeed greater than 8%. Note that, using simulation, we can derive some strong
conclusions about the relationship between maturity and return, even if we know nothing about
formal bond pricing theory.

Decision Models & Optimization

Page 10

Term II, 2016

Solutions to Self Study Problem Set #4


6. College Fund Investment Problem (OPTIONAL)
The spreadsheet collegeCB.xls contains a sample solution to this problem.

a) Set up a spreadsheet to simulate the two suggested investment plans over the 72-month period.
Plot the value of each strategy over time for one simulation trial. What was the total value of each
strategy after six years? Did either of the strategies reach the target?
Let Vt denote the value of the S&P 500 and Wt denote the value of the T-bill fund at the end of
month t. Denote the 261 historical returns by rt and st for the S&P 500 and T-bills, respectively.
Suppose that j is a random number between 1 and 261. Let C denote the total monthly contribution.
(Here, c = 200 the $200 per month that your neighbor has budgeted to invest.) Finally, let a
denote the fraction invested in stocks and 1-a the fraction invested in T-bills. These equations
describe the dollar values of Vt and Wt :

Vt = Vt - 1( 1 + rj) + a C
Wt = Wt - 1(1 + sj) + (1 - a) C
The Vt equation says that the account value of the S&P 500 fund account is given by the value in
the previous month (Vt - 1) times the random S&P return (1 + rj) plus the contribution of the fraction of
the $200 invested in stocks (aC). The S&P500 fund starts with V0 = aC at the end of month 0 (i.e.,
at the beginning of the problem.) The T-bill fund starts with W0 =(1 - a)C at the same time. The total
value of both funds at the end of the six years (72 months) is V72 + W72.
A partial view of the spreadsheet COLLEGECB.XLS is given in the figure below. A random month-j
from 1 to 261 is given as an ASSUMPTION in cell G14. The value of V1 is determined in cell H14
with the formula:
= H13*(1+VLOOKUP($G14,$A$9:$D$269,3))+H$13
= V0 * ( 1 + r1)
+
aC
The VLOOKUP function is used to find the historical return rj. The formula in cell H14 is copied to
the remainder of column H. (This assumes that Scott makes a final contribution at the end of month
72. If not, the last formula in cell H85 can be changed to remove the term H$13).
A similar formula is used to determine the value W1. In cell I14 the formula for W1 is
=I13*(1+VLOOKUP($G14,$A$9:$D$269,4))+I$13
= W0 * ( 1 + s1)
+ (1 - a) C
The same random month (from cell G11) is used to pick the S&P return and T-bill return.
Columns J and K compute the results for the second (20%:80%) allocation strategy. The total
ending values are recorded in cells H7 and J7. For our first trial, both strategies reached the
$17,500 target: strategy 1 (20% stock) finished with $19,045 while strategy 2 (80% stock) ended
with $21,595. A graph of the fund values over time is given in the graph below.

Decision Models & Optimization

Page 11

Term II, 2016

Solutions to Self Study Problem Set #4


A
B
C
D
E
1 COLLEGE1.XLS
2 OPIM 621: Decision Models & Uncertainty
3 Self Study #4 -- Problem 5 (OPTIONAL)
4
5
Historical Return Table
6
S&P 500
T-bill
7
Month
Yr/Mo Total Ret Total Ret
8
9
1
7001
-7.43%
0.60%
10
2
7002
5.86%
0.62%
11
3
7003
0.30%
0.57%
12
4
7004
-8.89%
0.50%
13
5
7005
-5.47%
0.53%
14
6
7006
-4.82%
0.58%
15
7
7007
7.52%
0.52%
16
8
7008
5.09%
0.53%
17
9
7009
3.47%
0.54%
18
10
7010
-0.97%
0.46%
19
11
7011
5.36%
0.46%
20
12
7012
5.84%
0.42%
21
13
7101
4.19%
0.38%
22
14
7102
1.41%
0.33%

Simulation parameters:
Monthly contribution
200
Fraction in S&P 500
0.2
Fraction in T-bills
0.8
Allocation 1 (forecast)
Total ending value:
17893

Month
0
1
2
3
4
5
6
7
8
9

Assumption
Random
month
187
246
121
115
51
13
134
70
147

200
0.8
0.2
Allocation 2 (forecast)
17183

S&P 500

T-bill

S&P 500

T-bill

Fund
40
80.6
128.5
171.0
218.0
258.4
313.5
360.0
388.2
425.8

Fund
160
320.9
483.1
647.7
812.9
977.6
1141.8
1314.0
1480.9
1655.4

Fund
160
322.5
514.0
683.9
871.9
1033.5
1253.9
1439.9
1552.7
1703.4

Fund
40
80.2
120.8
161.9
203.2
244.4
285.4
328.5
370.2
413.9

Total Fund Value vs. Time


30000

25000

20000

15000

10000

5000

Month
20% Stock

Decision Models & Optimization

Page 12

80% Stock

Term II, 2016

Solutions to Self Study Problem Set #4


b) Simulate 200 trials of the two strategies over the 72-month period using 1001 as the initial random
number seed. Create a histogram of the final fund values. Based on your simulation results, which
of the two strategies would you recommend? Why?
The summary statistics provided by Crystal Ball are displayed below. The average final fund value
was approximately $18,894 under strategy 1, and approximately $20,598 under strategy 2. Clearly,
investing more in stocks gives a higher average return. However, the risk is also higher, as
indicated by the standard deviations $1,012 and $4,061 for strategies 1 and 2, respectively.
Allocation 1

Fraction in S&P 500


Fraction in T-bills
Trials
Mean
Median
Mode
Standard Deviation
Variance
Skewness
Kurtosis
Coeff. of Variability
Range
Minimum
Range Maximum
Range
Width
Mean Std.
Error

0.2
0.8

Allocation 2
0.8
0.2

200
18894
18827
--1012
1023302
0.37
2.88
0.05
16809

200
20598
20232
--4061
16494032
0.41
2.91
0.20
11987

21819
5010

32186
20199

71.53

287.18

Crystal Ball generates histograms of each strategy for free. For clarity, we combined this
information by creating a customized histograms using the extracted forecast data. The customized
histogram is shown on the next page. Here it is easy to see that Strategy 1 peaks earlier and is
less spread out. The riskier strategy using 80% stock peaks later, with a wider dispersion.
Scott Jansen is most interested in hitting his target number of $17,500. The likelihood of reaching
the target under each strategy can easily be computed from the extracted data using an IF
statement. For example, in the worksheet Fund Value 500 trials (within spreadsheet
collegeCB.xls) we use the expression =IF(G11>17500,1,0) where G11 contains the final value for
trial 1 using strategy 1. This records a 1 if the final value exceeds the target, and 0 otherwise. The
probability of reaching the target can be estimated by averaging these values over all 500 trials.
Using this approach, the probability that the college funds value after six years exceeds the target
level of $17,500 under strategy 1 (80% bonds, 20% stock) is approximately 92.5%. Under strategy
2, this probability is roughly 75%. The 90% confidence intervals are 92.5% +/- 3.1% and 75.0% +/Decision Models & Optimization

Page 13

Term II, 2016

Solutions to Self Study Problem Set #4


5.1%, respectively. It appears that strategy 1, despite its lower return, has a better chance of
reaching the target.

Histogram of Final Fund Values


45
40
35
30
25
20
15
10
5
0
13

15

17

19

21

23

25

Final College Fund Value (in $000)

27
20% Stock

29
80% Stock

c) Suppose that your neighbor needs $19,500 to pay for the first year's tuition, rather than $17,500.
Based on the same simulation results, which of the two strategies would you recommend now?
Why?
Another way that the difference in the strategies can be viewed is with a plot of cumulative
frequencies, as shown in the figure below. For a given target value, one strategy is preferred to
another if the probability for exceeding the target is greater for the preferred strategy. The vertical
axis gives the probability of the fund value not exceeding the target. For a given target value, the
strategy with the lower cumulative frequency is preferred. The lines corresponding to the two
strategies cross at around $18,300. For targets less than $18,300, the conservative 80% bond
portfolio is preferred; for higher targets, the more risky 80% stock portfolio is preferred. Thus, for the
new target of $19,500, the stock portfolio is preferred.
With the higher target, our simulation indicated that strategy 1 reaches the target only 23% (+/5.1%). Strategy 2 reaches the target 51% (+- 5.7%). For the higher target, strategy 2 appears to be
better.

Decision Models & Optimization

Page 14

Term II, 2016

Solutions to Self Study Problem Set #4

Cumulative Frequencies of Fund Values


250

200

150

100

50

0
13

15

17

19

21

23

25

27

29

Final College Fund Value (in $000)


20% Stock

80% Stock

d) What other real-world factors might be important to consider in designing the simulation and making
a recommendation?
Two important factors not explicitly modeled are inflation and taxes. If inflation data were available,
the simulation could have been redesigned to reflect real returns instead of nominal returns. If taxes
become due in every year, for example, this greatly reduces the compounding of the initial
investment over long time periods.

Decision Models & Optimization

Page 15

Term II, 2016

You might also like