QUANTITATIVE DECISION MAKING

Dr. P.L. van den Berg & M. Pourakbar

Lecture 1: Decision making under certainty 2

A. What is quantitate decision making 2
B. Developing a decision table 2
C. Decision making under uncertainty 4
D. Decision making under risk 6
E. Sensitivity analysis for decision making under uncertainty 7
Lecture 2: Decision trees 9
A. Decision trees 9
B. Conditional probabilities 10
C. Value of imperfect information 11
D. Utility 12
Lecture 3: Simulation 14
A. Introduction 14
B. Monte Carlo Simulation 15
C. Running a simulation 17
Lecture 4: Linear programming 20
A. Introduction to linear programming 20
B. Formulating linear programs + graphical interpretation 20
Lecture 5: Linear programming - Sensitivity analysis 21
A. Impact of changing coefficients of the objective function 21
B. Impact of changing the right hand side of the constraints 22
Lecture 6: Integer linear program 23
A. Introduction to integer linear programming 23
B. Binary integer programming (BIP) 24
Lecture 1: Decision making under certainty

A. What is quantitate decision making

❖ Mission statement of the course: “We provide students with the methods and tools to make better, more
responsible, decisions in a complex and dynamic world!”
❖ In this course, quantitative methods will be used to support and improve the decision-making process.
❖ Decision = choice between different alternatives, in which the outcomes can be evaluated using different
criteria.
❖ To improve the quality of a decision, systematic methods are necessary. Our intuition is not always reliable.
❖ What makes a decision difficult?
- A lot of options
- Contradicting objectives
- Uncertainty about the outcome
❖ What makes a good decision?
- Different ways of approaching this:
• Evaluating the outcome of the decision ➔ good outcome = good decision
• Evaluating the process ➔ given the info that I had at the time, did I make right decision?
- You have to look at both to make a good decision
- Outcome bias: a bias in which people evaluate the quality of a decision solely based on the outcome.
- Best outcome is not always the result of the best decision

B. Developing a decision table

❖ Decision making under uncertainty appears in many situations in business

- Financial uncertainty
• When starting a new startup ➔ uncertainty in future success or demand for your product
• There is uncertainty on exchange contracts ➔ buy insurance?
• Long-term contracts or short-term auctions?
- Operational
• Amount of stock to hold?
• How much capacity should be reserved for premium customers?
• How much buffer time should be included in delivery routes?
❖ The news vendor problem
- This is a situation where you have o make a decision on the quantity to order when you do not
know the demand
- It is a basic inventory decision: costs of ‘too much’ vs. costs of ‘too little’.
- Limited shelf life: when buying (sporting)clothes, production of ice-skates, making bread, planning
production capacity…

2 of 25 2019 - 2020
❖ Decision problem: structure
1. Possible decisions (or strategies)
2. Possible (future) scenarios
3. Decision table = evaluation of possible decisions in different future scenarios
❖ Airline example: An airline has to decide on the number of flights a day between to destinations before the
start of the season. The aircraft of this airline have 200 seats and there can be 1 or 2 flights a day. The ticket-
price for this flight is €200 and the costs per seat are €100. When there is no seat available for a customer,
the airline looses ‘goodwill’ (costs € 80). The airline expects a demand between 100 and 400 per day. The
airline does not fluctuate their prices to steer demand and unsold seats have no value.
- Possible decisions
• 1 flight a day
• 2 flights a day
- Potential future scenarios
• Demand = 100, 200, 300, 400 (can be others but we discretize)
- Parameters:
• Cost per seat = €100
• Ticket-price = €200
• Penalty lost sales = €80
- The resulting decision table is as follows

- How many seats will we sell when we have a capacity of 200 and a demand of 300?
• We cannot sell more than capacity
• We cannot sell more than demand
• Hence: Sales = MIN(Capacity, Demand) = MIN(200,300) = 200

3 of 25 2019 - 2020
 - How much will the out-of-stock penalty be when the capacity is 200 and the demand is 300?
• If demand > capacity, then out-of-stock = demand – capacity
• But, out-of-stock cannot be lower than 0!
• Out-of-stock penalty = MAX(Demand – Capacity, 0) = MAX(300-200,0) = 100

❖ There are three decision types

- Type 1: under certainty = decision maker knows with certainty the consequences of each decision.
- Type 2: under uncertainty = decision maker knows nothing about the probability of each possible
outcome.
- Type 3: under risk = decision maker knows the probability distribution of the different outcomes.

C. Decision making under uncertainty

❖ Under uncertainty, there are 5 decision criteria

- MaxiMax (Optimistic)
- MaxiMin (Pessimistic)
- Criterion of realism (Hurwicz)
- Equal probability (Laplace)
- MiniMax regret

❖ MaxiMax criterion (Best-case)

- For each decision, determine the
best outcome (max), circled in
red
- Then select the decision with the
best outcome (max)

4 of 25 2019 - 2020
❖ MaxiMin criterion (Worst-case)
- For each decision, determine the
worst outcome (min) ➔ look for
the best worst case.
- Then select the decision with the
best outcome (max)

❖ Hurwicz criterion
- It is the weighted average
between optimism and pessimism
- Select a coefficient α between 0
and 1
• 1 ➔ 100% optimistic
• 0 ➔ 100% pessimistic
- Calculate the weighted average
for each alternative (decision)
- Select the alternative with the
highest value
- Weighted average = α(Best scenario) + (1 – α)(Worst scenario)
- Two special cases of this criterion
• If α =1 ➔ 100% optimistic ➔ MaxiMax criterion
• If α =0 ➔ 100% pessimistic ➔ MaxiMin criterion

❖ Laplace criterion (equally likely)

- For each alternative, determine the average outcome
- Select the best alternative (max)

❖ MiniMax regret
- Here we compute the “Opportunity loss” or “regret” = difference between the actual outcome and
the maximum possible outcome.
- We need to create an ‘opportunity loss’ table
by subtracting the outcomes of each column
from the best outcome (max) in that column.
- Then for each decision that can be made, we
deter mine the maximum regret per
alternative.
- Select the alternative with the lowest (min)
maximum (max) regret.

5 of 25 2019 - 2020
 D. Decision making under risk

1) The expected monetary value

❖ The decision used in this case is the

Expected monetary value, the weighted
average of the expected profit.
❖ This is about calculating the expected value.
The expected value is calculated as:
E(X ) = x1 p1 + x 2 p2 + . . . + xn pn .
❖ If in the airplane example, we introduce the
following probability distribution, the EMV
for each alternative is calculated as:

❖ With this criterion, we can also evaluate what the value is of better demand information.
- Expected Value with Perfect Information (EVwPI) ➔ expected profit that can be achieved knowing
demand in advance
• The average long-term outcome when having perfect information (prior knowledge).
- Expected value of perfect information (EVPI) ➔ value of perfect prior knowledge (the amount we
would be willing to pay for information about the future).
EVPI = EVwPI – Maximum EMV
❖ In the airplane example, a consultant claims to be able to predict the market perfectly. The costs of hiring
him are €5,000. Does it make sense to hire him ➔ is EVPI > €5,000.
- We first take the decision table
- For each demand scenario, we compute the outcome if we had had the information in advance
- For a demand of 100, we would have chosen a capacity of 200
- Same holds for demand of 200
- And so on…

6 of 25 2019 - 2020
 - Finally, we calculate the EMV
with perfect information.
- The maximum EMV without the
extra information was €11,200.
The EVPI is €18,000 – € 11,200
= €6,800. ➔ YES should hire
the consultant

2) Expected opportunity loss

❖ Expected Opportunity Loss (EOL): the weighted average of the expected regret.
❖ We are looking to minimize the expected regret
❖ Some observations:
- EOL criterion results in the same
decision as the EMV criterion
- Minimum EOL is equal to the
EVPI.
❖ Interesting parallel: min(EOL) = EVPI

E. Sensitivity analysis for decision making under uncertainty

❖ In a sensitivity analysis you investigate to what extent the situation would fluctuate if the variables changed.
In other words, how much will the outcomes vary when the input data is changed?
❖ Sensitivity analysis = will the decision change when the input variables are changed?
- Very important when the input data is unreliable or variable
❖ Robust model = decision does not change when the input parameters vary a lot. Robust models are
insensitive to variation/mistakes in input data
❖ Example: A company considers to enter a new market. To do this, the company can either build a big or
small factory. Furthermore, the company can choose not to enter the market and thus not build a factory.
The market can be positive or negative. The table below displays the profit in each scenario.

7 of 25 2019 - 2020
- Let P be the probability of a positive market
• EMV(big) = 200,000P – 180,000(1 – P) = 380,000P – 180,000
• EMV(small) = 100,000P – 20,000(1 – P)
• EMV(do nothing) = 0P + 0(1 – P) = 0
- We now compute the expected value for different values of P

8 of 25 2019 - 2020
Lecture 2: Decision trees

A. Decision trees

❖ Decision tree = way to graphically represent a decision table. Especially useful when making sequential
decisions.
❖ Two types of decision
- Decisions (square) ➔ up to manager’s choice
- Events (circles) ➔ not up to manager’s choice
❖ Example: we are entering a market,
- Three approaches to do so
• Aggressive
• Neutral
• Careful
- Future scenarios of the market
• Strong
• Weak
- The first decision is about
which approach to take
- After adding this decision,
we can add the different
events nodes, leading to the following decision tree:
- For each node, we can calculate the EMV
• EMV(agressive) = 30 × 0.45 + (−8) × 0.55
= 9.1
- So on for every possibility…

❖ Value of perfect information

- Expected value with perfect information (EVwPI) =
average long term outcome in the case of perfect
information (prior knowledge).
- Expected value perfect information = value of
perfect prior knowledge (the amount we would be
willing to pay for information on the future)
- EVPI = EVwPI – Maximum EMV
- To calculate the EVwPI, we swap the tree around ➔
- In this case, EVwPI is 21.75 and the best EMV is 12.85. Hence the value of perfect information is 8.9

9 of 25 2019 - 2020
❖ Now, suppose research does not provide perfect information, but we know the reliability of the information,
i.e. the probability of correct and wrong predictions
- Positive while the market is weak
- Negative while the market is strong
- Positive while the market is strong
- Negative while the market.
❖ What is the value of imperfect information?

B. Conditional probabilities

❖ Example: a security scan for a security system is active.

- Notation:
• P(A) = probability that alarm will ring
• P(B) = probability that there is a bomb
• P(A∩B) = probability of outcome A and B
P(A ∩ B) = P(A) × P(B | A)
• P(A|B) = probability of outcome A given B
P(A a n d B)
P(A | B) =
P(B)
- The system has the following reliability
• 99% of an alarm in case of a bomb P(A|B)
• 2% probability of an alarm in case of no bomb P(A|B’)
• 0.01% probability of a bomb P(B) = 0.01%
- What is the probe of a bomb given that there is an alarm
P(B ∩ A) P(B) × P(A | B)
P(B | A) = = = 0.0049
P(A) P(B′ ∩ A) + P(B ∩ A)
❖ Now suppose we have a market research that does not provide perfect information, but we know the
reliability of the information, but we know the reliability.
- Positive while the market is weak
- Negative while the market is strong
- Positive while the market is strong
- Negative while the market.
❖ Two types of predictions
- False positive
- False negative

10 of 25 2019 - 2020
❖ What if we know the probability that the market research is reliable:
- P(P|S) = 0.6 ( ✔ )
- P(N|S) = 0.4 (false negative)
- P(P|W) = 0.3 (false positive)
- P(N|W) = 0.7
- Plus: P(S) = 0.45 and P(W) = 0.55
❖ To determine these probabilities into probabilities
of the type P(S|P) and P(S|N)? Make a table such
as this one:

❖ Now what about the probability of S given P?

P(P ∩ S ) 0.27
P(S | P ) = = = 0.621
P(B) 0.435
❖ This is Baye’s theorem
P(S ) × P(P | S )
P(S | P ) =
P(S ) × P(P | S ) + P(W ) × P(P | W )

C. Value of imperfect information

❖ Market research presents extra (not fully trustworthy)

information on the market. There is a probability that
even thought the market research is positive (strong)
or that the research be negative and the market be
strong. The forecast is not fully accurate.
❖ What is the value of this information? This can be
determined by looking at decision trees.
❖ Probabilities are given:
- P(Strong market | Positive) = 0.621
- P(Weak market | Positive) = 0.379
- P(Strong market | Negative) = 0.319
- P(Weak market | Negative) = 0.621

11 of 25 2019 - 2020
❖ Also knowing that
- P(Negative) = 0.565
- P(Positive) = 0.435
❖ Should we conduct the market research

❖ Expected value of sample information = added value of extra information (the amount we would like to
pay for the research). It can have a value of 0.
❖ Expected value with sample information (EVwSI) = the average long term outcome with sample
information
EVSI = EVwSI – Maximum EMV
❖ Efficiency of sample information = a measure for the quality of sample information
EVSI
Efficiency of sample information = × 100
EVPI

D. Utility

❖ Expected monetary value (EMV) is not always a good indicator of the true value of an outcome of a
decision. In many situations, decision makers steer away from EMV.
❖ But decision makers are not immune for risk, they are not always risk neutral.
❖ The actual value of a decision maker is what we call the ‘utility’. A rational decision maker wants to maximise
his personal utility.
❖ Example: in a bowl there are
• 99 white balls for which you earn nothing
• 1 black call for which you earn €100.
- How much are you willing to pay?
- The answer will say whether you are a risk avoider or a risk seeker.
12 of 25 2019 - 2020
 - The EMV of this game is +€1.
• If one is willing to pay less than €1 ➔ risk avoider
• If one is willing to pay exactly €1 ➔ risk indifference
• If one is willing to pay more than €1 ➔ risk seeker.
❖ The first thing to do is to detmeine the risk sensitivity
- New game: Probability of winning 1, probability 1–p of losing 1
- Risk neutral decision maker:

- Risk avoiding decision maker

- Risk-seeking decision maker:

13 of 25 2019 - 2020
Lecture 3: Simulation

A. Introduction

❖ Example: you are hungry and you want as much pizza as possible: which of the two options will you choose?
[both pizza’s have the same thickness]
1. One pizza with a diameter of 2 cm and one with a diameter of 18 cm
2. Two pizzas with a diameter of 12 cm.The way to go about this could involve calculating the surface.
❖ Flaw of averages = if only considering the average input to determine the output, the outcome is wrong.
You should rather consider the average outcome.
❖ Simulation
- Imitate actual processes using a computer, sometimes using statistical probabilities. Especially in case
of complex system simulation can be faster, cheaper or safer.
- Simulation is one of the most widely used quantitative tools for analysis.
- Relatively simple to implement with current software
- Can be used to analyze large complex systems
- A lot easier than analytically (mathematically) analyze a system
❖ Pros and cons of simulation

Pros Cons
✓ Good way to test solutions without actually x Difficult and expensive to accurately simulate a
implementing changes. complex system
✓ Simulation can serve as a sensitivity analysis x Simulation does not present an optimal solution,
✓ Large and difficult systems can be analyzed but it is more like trial-and-error
✓ Flexible enough, not a lot of assumptions needed x Every simulation is unique. Solutions are not
✓ What-if analysis is possible transferable to other problems

Example decision problem:

- Data:
• Expected demand = 30
• Inventory = 25
• Price = 10€
• Total revenue = min(inventory; demand) × price
- What is the expected nonparty value? ➔ this cannot be calculated because we need a probability
distribution.

14 of 25 2019 - 2020
 - Now suppose we have access to the probability distribution

- One thing that we notice here is that Price × expected demand ≠ expected revenue ➔ flaw of
averages

❖ Building a simulation model consists of 7 steps

(1) Define the problem
(2) Introduce the variables associated with the problem
• Random variables ➔ no influence on those
• Decision variables ➔ variables whose value you can determine
(3) Construct a simulation model
(4) Specify input variables
(5) Run the simulation experiment
(6) Analyze the results
(7) Decide what course of action to take

B. Monte Carlo Simulation

❖ The principle of Monte Carlo Simulation is that a stochastic variable will be ‘simulated’ by a ‘randomly’
drawn number from a specific distribution.
- Often this simulation is part of a bigger simulation
❖ Applications of stochastic variables
- Number of students in the lecture
- Service time of a mechanic
- Time between the arrival of two customers
- Number of absent employees
❖ In Monte Carlo Simulation we use random numbers ➔ numbers that follow a certain distribution for which
there is no pattern.
❖ How to generate random numbers?
- Throwing a die
- Flipping a coin
- Computer based

15 of 25 2019 - 2020
❖ Pseudo-random numbers are the only ones that can be generated
- “Anyone who considers arithmetical methods of producing random digits is, of course, in a state of
sin”
- Pseudo-random numbers are generated by a formula
❖ Monte Carlo simulation: idea
- Pair a (collection of) random numbers to the values of the variables
- Pair the numbers in a way that corresponds with reality
- Generate a random number and you are performing a simulation
❖ 4 steps in Monte Carlo Simulation
(1) Establish probability distributions for input variables
(2) Construct a cumulative probability distribution for each variable in step 1
(3) Establishing an interval of random numbers for each variable
(4) Generate random numbers and simulate a probability distribution.

Example: from before, we again have the probability of demand

of various quantities

We can then determine the cumulative distribution ➔

Simulate with random numbers:
- 0.715 ➔ demand of 40.
- How about 0.213 ➔ demand of 20.

20 10 30 40 20

❖ Simulation techniques can be used with continuous distribution

16 of 25 2019 - 2020
 C. Running a simulation

Example: company sells products of which the demand is unknown. Furthermore, the lead time of the delivery
is also unknown. The owner wants to determine the expected costs (order costs, inventory costs, out-of-stock
costs) for a given order quantity and order point. In this model
- Uncertainty (cannot be influenced)
• Demand
• Lead time
- Decisions
• Order quantity
• Order moment
- Goal
• Minimize costs
The following flowchart represents the situation

In this case, we set:

- Order quantity = 10
- Order moment = 5

Step 1: probability distributions

17 of 25 2019 - 2020
Step 2: Determine cumulative probabilities

Step 3: select intervals

Step 5: run the Monte Carlo simulation

Step 5: Analyze the results

❖ Notes:
- Simulating 10 days is way too short
- Easier to simulate a lot of days
- More simulation will give better overview

18 of 25 2019 - 2020
❖ Verification and validation to check model
- Verification: is the model implemented correctly
• Did we build the model right?
- Validation: does the model correspond with reality?
• Did we build the right model
- If we know the actual outcomes for a set of inputs, we can compare this to the outcomes of the
model for those same inputs.

19 of 25 2019 - 2020
Lecture 4: Linear programming

A. Introduction to linear programming

❖ Linear programming = method used to assist decision-making by using quantifiable goals.

- This is a general concept related to optimization subject to constraints.
- This is an extremely powerful due to the smaller usage of mathematics and doesn’t lead to high
costs.
❖ The combination of this powerfulness and generality of concept leads to a wide number of applications
❖ Characteristics of linear programming:
- Linear programming model tries to either maximize or minimize a certain linear objective function
subject to a number of constraints.
- Model consists of:
• Decision variables
• Objective (minimize/maximize)
• A number of constraints
❖ Assumptions of liar programming:
- Values of parameters are known with certainty and will not change in the relevant time interval
- Value of the objective and the constraints will change proportionally with the decision variables
(linear)
- Decision variables can have any value within the feasible region (not necessarily integers)

B. Formulating linear programs + graphical interpretation

See notes on the book. It’s the same thing with a (simple) example

20 of 25 2019 - 2020
Lecture 5: Linear programming - Sensitivity analysis

A. Impact of changing coefficients of the objective function

1) Graphical interpretation

The case that we will look at:

❖ The what if analysis that we want to conduct is

- What happen if the profit per unit of one of the
product changes?
- What happens if the production capacity of one of the
factories changes?
❖ Experimenting with different values of the price and obtain the
following graph
❖ At certain break points, the slope changes. What are they? We visualize
them looking at a table:
❖ They correspond to the new optimal production quantities.

2) In excel

❖ Conducting a sensitivity analysis in Excel yields the kind of table

on the right
- Allowable increase: how much higher the price can be
increased without changing optimal solution
- Allowable decrease is how much more the price can
be decreased without changing the optimal solution
❖ The interval “allowable range” within which the optimal
solution does not change, can be calculated by:
[Objective coefficient – Allowable decrease, Objective coefficient + Allowable increase]

21 of 25 2019 - 2020
 B. Impact of changing the right hand side of the constraints

1) Graphical interpretation

❖ Shadow price: rate of change in the objective function value with a unit change in the right hand side of a
functional constraint
❖ For every functional constraint, we can find the associated shadow price.
❖ It represents the economic value of a scarce resource for the company
- How much should the company be willing to pay for one more unit of that resource?

❖ With the allowable range, the optimal function value changes proportionally with the shadow price
- If the changed constraint is binding, then the optimal solution and objective function value change
proportionally with the shadow price.
- If the changed constraint is not binding and the change is within the allowable range, then the
optimal solution and the optimal function value remain the same.
❖ Outside the allowable range, the structure of the solution changes ➔ new binding constraints (new shadow
price).

22 of 25 2019 - 2020
Lecture 6: Integer linear program

A. Introduction to integer linear programming

❖ We do not want to solve integer linear programs ➔ we just want to formulate them
❖ In general, linear programming solutions are not integers (even if parameter are set as integers), the optimal
solution is most of the time non integer
❖ When is a non integer solution OK?
- Decision variables are in nature divisible units (e.g. kg, $…)
- Decision variables are rates (e.g. units per week)
❖ However, sometimes, rounding is acceptable ➔ when numbers are big
❖ But sometimes we cannot round!
- When numbers are small
- Yes-or-no decisions
❖ The main relationship between LP and IP is graphically
visible
- In LP, the feasible region is an area
- In IP, the feasible region is not an area but a
discrete set of points
- The IP region is a subset of the LP region
❖ Definitions
- Relaxed ILP = the problem without the integrity-requirements
- LP relaxation = a method used to solve large ILP problems by relaxing the relevant problem to an
LP which can be solved much more easily.
❖ The optimal objective function value of the relaxed problem is
- An upper bound for ILP in case of a max problem
- A lower bound for ILP in case of a min problem
❖ Example: Production planning, 2 products, 2 constraints
- Max: 100A + 121B
- Subject to
A + 5B ≤ 24
5A + 6B ≤ 50
A, B ≥ 0
- The LP solution is: A = 5.58 ; B = 3.68
- What is the best integer solution?

23 of 25 2019 - 2020
 B. Binary integer programming (BIP)

1) Basics

❖ This is a special case of integer programming because they are variables that can get either 0 or 1
- Pure BIP: all decision variables are binary
- Mixed BIP: not all but some of the variables are binary
❖ Used to model
- Yes/No decisions
- Fixed costs
- Selection among limited number of alternatives
- Dependency conditions
❖ Example: Pharma Corp is looking for a new medication
- 5 potential R&D projects
• Project up
• Project stable
• Project choice
• Project hope
• Project release
- Available budget: $1.2
- Investment needed, success rate, and revenue in case of success are given for each project…

- Question: Which project(s) must be selected for R&D to maximize expected profit?
• Define xi = 1 if project i is selected
• Maximize: P = 300x1 + 120x 2 + 170x3 + 100x4 + 70x5
• Subject to:
R&D budget: 400x1 + 300x 2 + 600x3 + 500x4 + 200x5 ≤ 1,200
and xi ∈ {0,1} for i = 1,2,3,4,5

24 of 25 2019 - 2020
 2) Modeling additional constraints

❖ Now the previous company cannot select more than 4 projects

- New constraint: x1 + x 2 + x3 + x4 + x5 ≤ 4
- The company will select only two projects: x1 + x 2 + x3 + x4 + x5 = 2
❖ Dependency conditions
- Project 1 can only be selected if project 2 is selected: x1 ≤ x 2 (or x1 − x 2 ≤ 0)
- Project 1 and project 2 could be done if they are both selected: x1 = x 2 (or x1 − x 2 = 0)
❖ Fixed costs: required to produce new items:
- Y1 is the production quantity of product 1. This product can only be produced if project 1 is
selected. The maximum production quantity of this product is 1500.
Y1 ≤ 1500x1 (or Y1 − 1500x1 ≤ 0)
- Y2 is the productions quantity of product 2. This product can only be produced if project 1 is NOT
selected
Y2 ≤ 1500 ⋅ (1 − x1) (or Y2 − 1500 ⋅ (1 − x1) ≤ 0)

25 of 25 2019 - 2020