BC2410/2411, Prescriptive Analytics

From Data to Decisions

Lecture 12
Schedule
Schedule
Optimization under Uncertainty

 Parameters of the model c, A, and b may be uncertain when decision

x is made
 How would one describe uncertainty?
 How would one evaluate the “best” solution under uncertainty?
Facets of Uncertainty
 Risk: Uncertainty with known frequency of occurrence
 Ambiguity: Uncertainty without known frequency of occurrence
• Also known as Knightian uncertainty, due to economist Frank Knight (1885-
1972)
Facets of Uncertainty: Risk

 Uncertainty described via Random Variables

• Widely accepted method in math/statistics
• Characterized by
• Sample space (all possible outcomes, usually exponential or infinite)
• Distributions (probability of outcome)
• Optimize via Stochastic Programming
Decisions Under Risk $20M

0.5

A
0.5
B $0

$5 M

 A is preferred in decision tree.

• Why?
Bernoulli Distribution
Bernoulli Distribution
 1 trial

 Has exactly two outcomes : “success” or “failure”

 Success p, failure 1 – p

𝑿 = 1] = p
𝑿 = 0] = 1 – p

𝑿 = x] = px(1– p)1-x for x = 0,1

Bernoulli Distribution
 1 trial

 Has exactly two outcomes : “success” or “failure”

 Success p, failure 1 – p

𝑿 = x] = px(1– p)1-x for x = 0,1

Mean:

Variance:
Binomial Distribution
 n independent trials

 Each trial has exactly two outcomes : “success” or “failure”

 Each trial has same probability: success p, failure 1 – p

 ,𝑛
𝑛−𝑥 !

Binomial Distribution
We also say that X obeys a binomial distribution with
parameters n and p : Binomial(n, p) or B(n,p)
Probability
mass function
(pmf)
Binomial Distribution

Mean:

Variance:
Probability Mass Function (pmf)
Cumulative Distribution Function (cdf)
Poisson
Distribution
Poisson Distribution

It all starts with Mala Beef La Mian …

Poisson Distribution
Due to the outbreak of COVID-19, many students chose not to have
their lunch or dinner at South Canteen. The cook of Xiao Long Bao
didn’t know how many Mala Beef La Mian to prepare, so he collected
some lunch data for the first week:
# of MBLM sold
Mon 3
Tue 7
Wed 4
Thu 6
Fri 5

What is the probability of selling k MBLM during lunch time?

Poisson Distribution Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 Yes Yes Yes No

14:00

p: probability of selling one MBLM in each interval

Y: number of intervals that have 1 MBLM sold
Poisson Distribution Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 14:00

T
Poisson Distribution Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 14:00

T
Poisson Distribution Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 14:00

How to obtain the probability p?

Poisson Distribution Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 14:00

This is the probability of selling k Mala Beef La Mian within time T!!
Poisson Distribution
A random variable X is said to be a Poisson r.v. with parameter
λ (> 0) if it has the probability function

Mean:

Variance:
Poisson Distribution
Useful for modelling the number of occurrences of an event over a specified
interval of time or space.
Examples :
• Number of customer orders received in one hour
• Number of failures in a large computer system per month

Properties:
• Probability of an occurrence is the same for any two intervals of equal length.
• Occurrences in nonoverlapping intervals are independent of one another.
Poisson Distribution
 Examples and Applications:
– The number of phone calls arriving at a call centre per minute.
– The number of goals in sports involving two competing teams.
– The number of infant death per year.
– The number defaults found per newly TOP flat.
–…
Exponential
Distribution
Exponential Distribution

… …
T1 T2 T3 T4 T5

Ti: time interval between event i – 1 and event i

Y = time interval between selling two consecutive MBLMs

Exponential Distribution

… …
T1 T2 T3 T4 T5

If there is a unit of time that you sell nothing, then your probability is:
Exponential Distribution (Optional)

t
# of item sold every unit of 1
time
# of item sold half unit of
time
# of item sold a quarter unit
of time
Exponential Distribution
 Often arises as the distribution of amount of time until an event occurs.
 Useful for waiting line problems.

A random variable X is said to be an exponential r.v. with

rate parameter λ ( > 0) if it has the pdf
f ( x ) =e   x for x > 0
Exponential Distribution
Stochastic Programming
 An approach of optimization under risk
 Optimization that is based on a collection of uncertain outcomes
• Historical data
• Generated by Monte Carlo Simulation
Xiao Long Bao’s Problem
Due to the outbreak of COVID-19, many students chose not to have
their lunch or dinner at South Canteen. The cook of Xiao Long Bao
didn’t know how many Mala Beef La Mian to prepare, so he collected
some lunch data for the first week:
# of MBLM sold
Mon 3
Tue 7
Wed 4
Thu 6
Fri 5

How many MBLM to plan before the lunch time?

Xiao Long Bao’s Problem Mon
Tue
# of MBLM sold
3
7
Wed 4
Thu 6
Fri 5

11:00 14:00

T
A Newsvendor Problem
 Buy x items at cost c
 Items sold at price p
 Salvage price is zero
 Uncertain demand, hence profit is uncertain
Summary
A Newsvendor Problem
 Buy x items at cost c
 Items sold at price p
 Salvage price is zero
 Uncertain demand, hence profit is uncertain
A Newsvendor Problem
 Assume that the process can be repeated indefinitely under identical
conditions justify one to maximize expected profit
• Closed form solution available.
A Newsvendor Problem
 Assume that the process can be repeated indefinitely under identical
conditions justify one to maximize expected profit
• Closed form solution available.
 We can solve a stochastic programming problem by generating data
using Monte Carlo simulation.
Generate Demand
A Newsvendor Problem

 Generate samples using Monte Carlo approach.

 Observe that at instance k, the demand is .

 How much one can sell is or

A Newsvendor Problem
 Sampling Average Approximation (SAA) of the Newsvendor problem.
 Quality of the solution improves with the number of samples.
Multi Product Newsvendor Problem
 N items to sell.
 For each item i, unit cost is , selling price .
 Order quantity , i = 1, … , N.
 Assume salvage price is zero.
 Each item i incurs storage space . Total space available is C.
 K instances of demands available.
 Let denotes the demand of item i in instance k, k = 1, …, K.
 Close form solution will be difficult!
Multi Product Newsvendor Problem
Robust Newsvendor Problem
 For the Newsvendor problem, what if we do not know the demand
distribution?
Additional resources
 Advanced courses offered at NTU
• MH3400: Algorithms for the Real World
• MH3701: Basic Optimization
• MH4701: Mathematical Programming
• MH4702: Probabilistic Methods in OR
• MA4804: Optimization Theory & Applications
 Online resources
• Operations Analytics (Wharton, Coursera)
• Analytics for Decision Making Specialization (Minnesota, Coursera)
• Discrete Optimization (Melbourne, Coursera)
 Master program