1. Introduction to Operations Research
Statistical Concepts, Optimization, Heuristics, Simulation, and Forecasting
Judy Pastor
Steven Coy

2. Sum and Product Notation
A sum of a row or column of n numbers
Ex. X = (1, 3, 4, 9)
= 13  4  9 = 16
An iterated sum: Sums a matrix of numbers having n rows and j columns
A sequential product of n numbers
= 13  4  9 = 108

3. Convex and concave sets
Convex Set
The set of all the points that are bounded by this curve
Concave Set
The set of all points that are bound by this curve

4. Probability Concepts Experiment Sample outcome Sample space Event
A repeatable procedure
Has a well defined set of possible outcomes
Sample outcome
Potential result of an experiment, denoted e
Sample space
The set of all possible outcomes, denoted S
Event
A subset of the sample space corresponding to the definition of the event, denoted E

5. Probability of an Event
Probability is the degree of chance or likelihood that and event will occur in an experiment
Calculating the probability for a discrete or countable problem
1. Find the sum of possible outcomes that satisfy the definition of the event
2. Find the sum of the total number of possible outcomes
3. Divide the result in 1 by the result in 2
In mathematical notation
P(E) = e  {E} /  e  {S}
P(E)  P(S) = 1

6. P(FreqFlyer) = 1MM/5MM = .20 or 20%
Example
Experiment: Pick a pax from the passenger database
Sample Space: 5 million total pax; 1 million pax are
frequent fliers
Event: Passenger is a frequent flier
P(FreqFlyer) = 1MM/5MM = .20 or 20%
Note: P(S) = 1

7. Union and Intersection of Two Events
The sum of the sample outcomes of two or more events that are common to all of these events
A  B
Typically identified by the word “and” as in A and B
Union
The sum of the sample outcomes of two or more events
AB = A + B - A  B
Typically identified by the word “or” as in A or B
Probability of the Union of Two Events
P(AB) = P(A) + P( B) - P( A  B )

8. Example Experiment: Pick a pax at random from the passenger database
Sample Space: 5 million total pax; 1 million pax are
frequent fliers; 2.7 million pax are female;
600 thousand frequent fliers are female
Event: Passenger is a female or a frequent flier
P(Female) = 2.7MM/5MM = 54%
P(FreqFlyer) = 20%
P(Female and FreqFlyer) = 0.6MM/5MM = 12%
P(Female or FreqFlyer) = 54% + 20% - 12% = 62%

9. Conditional Probability
Conditional probability is the probability of an event, A, given that a related event, B, has already occurred
P(A|B) = P(A  B)/P(B)
Conditional probability effectively reduces the size of the sample space

10. Example
Experiment: Pick a passenger at random from the passenger database
Sample Space: 5 million total pax; 1 million pax are
frequent fliers; 2.7 million pax are female;
600 thousand frequent fliers are female
Event: Passenger is a frequent flier given that the pax is female
P(Female and FreqFlyer) = 0.6MM/5MM = 12%
P(Female) = 54%
P(FreqFlyer| Female) = 0.12/0.54 = 22.2%

11. Calculating Expected Value
Expectation is a long-run weighted average
Example
What is the expected return from running a revenue integrity process?
The process, which searches for duplicate bookings and expired TTLs, costs $0.1 for every record that the process scans.
For each duplicate reservation found, we receive $100 in incremental revenue and for each expired TTL, we receive $25.
We know that the long-run probability that a reservation will have a dupe is P(D) = 0.15% and that a TTL is expired is P(TE) = 0.1%
If we use the process to scan 100 K reservations, what is the expected return?
This requires an expected value computation
E(R) = P(D) * $100 + P(TE) * $25 - $0.10
E(R) = 0.15% * $ % * 25 - $0.10 = $0.175
E(R/100 K) = $0.05 * 100 K = $5000

12. Random Variables
Random Variables (RV) are characteristics or outcomes that vary from observation to observation
Independence of two RVs
two RVs are independent if the outcome of one does not effect the outcome of another => P(A|B) = P(A)
Correlation of two random variables
Two RVs are correlated if the knowledge of the outcome of one gives us an indication of the outcome of the other
Positive: X moves with Y
Negative: as X increases, Y decreases

13. Probability Distribution of a RV
Consider the unconstrained demand for a leisure class ticket on Flt 102:
Let’s compile the demand for each departure of Flt 102 for a full year and create a frequency histogram.
Notice that the histogram is mound-shaped and approximates a familiar bell-shaped curve.

14. Probability Distribution of a Continuous RV
The bell-shaped curve that we saw on the last slide is a normal density curve
Using this chart, we could argue that demand for this flight is normally distributed
Probability calculations with a normal distribution
Example: What is the probability that demand will be less than or equal to 35 pax?
First, we standardize the curve--transform it so that the area under the curve is equal to 1 (we use a z-transform)
We then find the area under the curve that satisfies the definition of our event (the interval 0 to 35)
The area under the curve from 0 to 35 = P(D  35)  .84

15. Calculating a Normal Probability
To find the probability, we find the interval on the horizontal axis and calculate the area under the curve corrsponding to that interval
P(D<X)=A

16. More About Distributions
Cumulative distribution
In our demand example, we found a probability for a single value of x
A cumulative distribution gives us the probability of D  x for any value of x
Cumulative Normal distribution

17. Truncated Normal Distribution
Demand processes that are described by a normal distribution are truncated at 0--demand cannot be negative

18. Some Common Distributions
Poisson: used to describe the arrival pattern (or process) of people to a system
For example, 5 people request a certain fair class product per hour
Mean = variance
Exponential: used to describe service times
If a process generates poisson arrivals, the generating process is exponential
Has special properties that makes it a favorite for queueing systems (waiting lines) and reliability
Uniform (a favorite, because it is easy): the probability is the same throughout
ExampleU(10, 20): P(x > 15 ) = P(10 < x < 15)

19. Central tendency Error measures Statistical Measures Mean = average
Median = middle value in a sorted list
Mode = largest value or highest portion of a probability density curve
Error measures
e = x - prediction of x
MAD(MAE): Mean absolute deviation (error):  |e| / n
MAPE: Mean absolute percentage error  |e| /x / n

20. More Measures
Variance: Measures the dispersion (spread) of the observations
Standard deviation: The square root of the variance
Coefficient of variation: The standard deviation divided by the mean--stated as a percentage
Skew: Measures the asymmetry of a distribution
mean > median: Positive or right-skewed
mean < median: Negative or left-skewed
Correlation: Statistical measure of the relationship between two variables from -1, perfect negative correlation to 1, perfect positive correlation

21. Truncated Normal Distribution
In RM, the tails of the normal demand curve typically extend beyond the capacity of the plane
This is why we use unconstraining (detruncation) algorithms to approximate the tail of the curve
Cap = 125

22. Operations Research
“application of mathematical techniques, models, and tools to a problem within a system to yield the optimal solution”
Phases of an OR Project
formulate the problem
develop math model to represent the system
solve and derive solution from model
test/validate model and solution
establish controls over the solution
put the solution to work

23. Linear Programs – A Major Tool of OR
Linear Programs (LPs) are a special type of mathematical model where all relationships between parts of the system being modeled can be represented linearly (a straight line).
Not always realistic, but we know how to solve LPs.
May need to approximate a relationship that is slightly non-linear with a linear one.
When to use: if a problem has too many dimensions and alternative solutions to evaluate all manually, use an LP to evaluate.

24. Linear Programs – A Major Tool of OR
LPs can evaluate thousands, millions, etc. of different alternatives to find the one that best meets the objective of the business problem.
Fleet Assignment Model - assign aircraft to flight legs to minimize cost and maximize revenue
Revenue Management - set bid prices to maximize revenue and/or minimize spill
Crew Scheduling - schedule crew members to minimize number of crew needed and maximize utilization

25. Linear Program Formulation
Understand the system and environment to which the problem belongs
Understand the problem and the objective to be achieved
State the model - clear idea of problem and what can and can not be included in the model
Collect Data - get data/parameters/constraints and boundaries of system and interrelationships
Determine decisions - define decision variables - what do we need the model to tell us?
Formulate and solve model

26. Example: RM Network LP
Problem - how many passengers of each itinerary and fare class should be accepted on each flight to achieve the maximum revenue for the flight network?
Statement - the model should tell us the above
Data - demand by itinerary/fare class, aircraft capacity, overbooking levels, expected revenue by itinerary/fare class
Decisions - how many passengers of each itinerary/fare class to accept on each flight leg

27. Example: RM Network LP Data Collection
Two Flights: SFO-IAH, IAH-AUS
Two fare classes: Y-high fare, Q-low fare
Three itineraries: SFOIAH, IAHAUS, SFOAUS
Six fares:
Flight capacity: SFO-IAH 124, IAH-AUS 94
No overbooking
Fares
Market Y Q
SFOIAH
400
300
IAHAUS
250
100
SFOAUS
450
320

28. Example: RM Network LP Data Collection
Demand
Market Y Q
SFOIAH 30 90
IAHAUS 50
SFOAUS 20

29. Example: RM Network LP Formulation Model
Data Definition:
F set of flights = {SFOIAH, IAHAUS}
f index of F (1,2)
CAPf capacity of flightf = {124, 94 }
I set of itineraries {SFOIAH, IAHAUS, SFOAUS}
i index of I (1,2,3)
IFf set of itineraries over flight f
IF1={SFOIAH,SFOAUS} IF2={IAHAUS,SFOAUS}
C set of classes {Y, Q}
c index of C (1,2)
DMDi,c demand for itinerary i and class c
FAREi,c fare for itinerary i and class c

30. Example: RM Network LP Formulation Model
Now that we have defined all the data that we know about the model, we now must define what we want the model to tell us.
Problem - how many passengers of each itinerary and fare class should be accepted on each flight to achieve the maximum revenue for the flight network?
Define decision variables:
Xi,c # pax accepted for itinerary i and class c
There are 3 itineraries and 2 classes so there are a total of 6 decision variables.

31. Example: RM Network LP Formulation Model
Many sets of values (collectively called solutions) for the six Xi,c variables exist which could satisfy the constraints (formulation coming) of aircraft capacity and maximum demand. These are “feasible” solutions.
Which solution do we want?
Problem - how many passengers of each itinerary and fare class should be accepted on each flight to achieve the maximum revenue for the flight network?
The feasible solution for this is “optimal”.

32. Example: RM Network LP Objective Function

33. Example: RM Network LP Obj. Function & Constraints
The Objective Function is an expression that defines the optimal solution, out of the many feasible solutions. We can either
MAXimize - usually used with revenue or profit or
MINimize - usually used with costs
Feasible solutions must satisfy the constraints of the problem. LPs are used to allocate scarce resources in the best possible manner. Constraints define the scarcity.
The scarcity in this problem involves a fixed number of seats and scarce high paying customers.

34. Example: RM Network LP Constraints

35. Example: RM Network LP Constraints
Rules for Constraints
must be a linear expression
decision variables can be summed together but not multiplied or divided by each other
have relational operators of =, <=, or >=
must be continuous
Constraints define the “feasible region” - all points within the feasible region satisfy the constraints.
The feasible region is convex.
The optimal solution lies at an extreme point of the feasible region.

36. Example: RM Network LP Cplex Input File
MAX
400 X_SFOIAH_Y X_SFOIAH_Q +
250 X_IAHAUS_Y X_IAHAUS_Q +
450 X_SFOAUS_Y X_SFOAUS_Q ST
CAPY_SFOIAH: X_SFOIAH_Y + X_SFOIAH_Q + X_SFOAUS_Y + X_SFOAUS_Q <= 124
CAPY_IAHAUS: X_SFOAUS_Y + X_SFOAUS_Q + X_IAHAUS_Y + X_IAHAUS_Q <=94
DMD_SFOIAH_Y: X_SFOIAH_Y <= 30
DMD_SFOIAH_Q: X_SFOIAH_Q <= 90
DMD_IAHAUS_Y: X_IAHAUS_Y <= 50
DMD_IAHAUS_Q: X_IAHAUS_Q <= 30
DMD_SFOAUS_Y: X_SFOAUS_Y <= 20
DMD_SFOAUS_Q: X_SFOAUS_Q <= 50
END

37. Example: RM Network LPSolution - Constraints
SECTION 1 - ROWS
NUMBER ROW AT ...ACTIVITY... SLACK ACTIVITY ..LOWER LIMIT. ..UPPER LIMIT. .DUAL ACTIVITY
1 obj BS NONE NONE
2 CAPY_SFOIAH UL NONE
3 CAPY_IAHAUS UL NONE
4 DMD_SFOIAH_Y UL NONE
5 DMD_SFOIAH_Q BS NONE
6 DMD_IAHAUS_Y UL NONE
7 DMD_IAHAUS_Q BS NONE
8 DMD_SFOAUS_Y UL NONE
9 DMD_SFOAUS_Q BS NONE
obj is the objective function value - total revenue from the small network of flights
CAPY_SFOIAH and CAPY_IAHAUS are the capacity constraints. Both are at UL - upper limit with activities of 124 and 94, respectively (i.e. both flight legs are full).
Dual Activity on each capacity constraint is also known as the Shadow Price of the flight. The SP of SFOIAH is 300 and the SP of IAHAUS is In RM terms, this means that the value of one more seat on SFOIAH is 300 and the value of one more seat on IAHAUS is Alternately, 300 and 100 also define the lowest fare that should be accepted on each leg.
DMD_{SFOIAH,IAHAUS}_{Y,Q} are the demand constraints. SFOIAH_Y, IAHAUS_Y, and SFOAUS_Y are at upper level (i.e. accept all Y passengers). Reject some/all of Q.

38. Example: RM Network LP Solution - Decision Variables
SECTION 2 - COLUMNS
NUMBER COLUMN..... AT ...ACTIVITY INPUT COST.. ..LOWER LIMIT. ..UPPER LIMIT. .REDUCED COST.
10 X_SFOIAH_Y BS NONE
11 X_SFOIAH_Q BS NONE
12 X_IAHAUS_Y BS NONE
13 X_IAHAUS_Q BS NONE
14 X_SFOAUS_Y BS NONE
15 X_SFOAUS_Q LL NONE
This section of the solution report shows the values for the decision variables at the optimal solution.
The LP tells us to accept 30 SFOIAH Y, 74 SFOIAH Q (reject 16), accept 50 IAHAUS Y, accept 24 IAHAUS Q (reject 6), accept 20 SFOAUS Y, and accept no SFOAUS Q.
Note that the LP cut off all SFOAUS Q booking requests because their fare of 320 is less than the sum of the shadow prices of the two flights ( = 400 > 320).

39. Example: RM Network LP Solving LPs
A problem that sounds small, like our example, can balloon out into many decision variables and constraints.
Computer software is available to solve linear programs.
Cost of programs depends on size of problems to be solved.
Excel has an Add-in to solve small LPs.
CPLEX is state of the art, but more expensive.
LPs with 100,000s row and columns can be solved.

40. Example: RM Network LP Solving LPs
The first method of solving LPs was invented during WWII by George Dantzig. The algorithm is called SIMPLEX. It is based on convexity theory and that the optimal solution will occur at an extreme point of the solution space
Newer state of the art algorithms are based on steepest descent gradient methods and are called “interior point” methods
Interior point methods can be extremely fast (much faster than SIMPLEX) for certain structures of problems

41. Degeneracy
When an LP has more than one unique way to reach an optimal objective function value, we say that the problem is “degenerate”
LP solvers can detect degeneracy but only report one solution
It would be nice to see all possible solutions
Different solvers can “land” on different solutions of a degenerate problem, depending on solution strategy
The RM Network problem is usually degenerate

42. Other Types of Linear Optimizations
MIP (Mixed Integer Programming)
is similar to LP but at least one decision variable is required to be a integer value
violates the LP rule that decision variables be continuous
is solved by “branch and bound” - solving a series of LPs that fix the integer decision variables to various integer values and comparing the resulting objective function values
is done in a smart way to avoid enumerating all possibilities
is useful, since you can not have .3 of an aircraft

43. Other Types of Linear Optimization
Network problem
is a special form of LP which turns out to be “naturally integer”
can be solved faster than an LP, using a special network optimization algorithm
is very restrictive on types of constraints that can be present in the problem
Shortest Path
finds the shortest path from the source (start) to sink (end) nodes, along connecting arcs, each having a cost associated with them
is used in many applications

44. Other Optimization Models
Quadratic Program
has a quadratic objective function with linear constraints
can be applied to revenue management, because it allows fare to rise with demand within a problem
price(OD) = 50 + [5*numpax(OD)]
max revenue = price * numpax

45. Other Optimization Models
Non-linear Program (NLP)
can have either non-linear objective function or non-linear constraints or both
feasible region is generally not convex
much more difficult to solve
but it is worth our time to learn to solve them since world is actually non-linear most of the time
some non-linear programs can be solved with LPs or MIPs using piecewise linear functions

46. Deterministic versus Stochastic
Two broad categories of optimization models exist
deterministic
parameters/data known with certainty
stochastic
parameters/data know with uncertainty
Deterministic models are easier to solve. Our RM LP is deterministic (we pretend we know the demand with certainty).
Stochastic model are difficult to solve. In reality, we know a distribution about our demand. We get around this in real life by re-optimizing.

47. Deterministic versus Stochastic
Deterministic optimization ignores risk of being wrong about parameter/data estimates.
No commercial software packages are currently available to do generalized, stochastic optimization.

48. Heuristics Definition - educated guess
When you use a heuristic to solve a problem, you have a gut feeling that it is a pretty good solution, but can not prove it mathematically
You can not prove that there is not a better solution out there
To qualify as an “optimal” solution, there must be a mathematical proof to say that no better solution exists

49. Types of Heuristics
“Greedy Algorithms” - also called “myopic” - nearsighted solutions.
Example: in our RM Network LP, the greedy solution would be to take the highest fare passengers possible on each leg, without looking at the consequences of doing so on the connecting leg. So the greedy solution is to take the SFOAUS Q passengers at a fare of $320. But the optimal solution looks at displacement and says do not take any SFOAUS Q passengers.

50. Heuristics
Combinatorial problems can grow exponentially when the number of decisions needed to be made grows linearly. Heuristics can be used in these cases to get a good solution in a reasonable amount of time.
TSP - Travelling Salesman Problem is a good example of this.
EMSR is a heuristic. It is provably optimal for two fare classes, but not more. However, it gives a good answer in a finite amount of time and takes probability into account.

51. Simulation
When a problem is too complicated to be put into an LP or a solvable non-linear optimization, one way to study the problem is to simulate it under different conditions.
PODS (Passenger O & D Simulation) is one example.
Simulation can tell us something about a set of parameters (i.e. total revenue, load factor), but does not point us in the direction of an improvement.

52. Forecasting Types of forecasting techniques used in RM:
pick-up
booking regression
exponential smoothing
Pick-up - adds average future bookings from historical observations to bookings on hand.
Booking Regression - computes best fit for history of bookings on hand (independent) to final booked (dependent)
final booked = a + b*(bookings on hand)

53. Forecasting
Exponential Smoothing - similar to pick-up except the average is weighted. The most recent historical observations are weighted most heavily, decreasing for earlier observations.
recursive relationship
Avg Pick Up = a * (pick upt-1) + (1-a)2* (pick upt-2) + ...
boils down to
Avg Pick Up = a * (pick upt-1) + (1-a) * (last fcst pick up)
Problem is how to estimate a. 0<=a<=1
How much weight should be on most recent observation?

54. Questions?