Chapter 1: Introduction to Operations Research

The Modelling Process

Mathematical decision-making model
- Variables → decisions
- Parameters → constants
- Measure of performance
- Constraints → restrictions

Step 1: Formulation of the model and data collection

- Defining the problem
- Developing a model
- Acquiring input data
Step 2: Solution development and validation of the model
- Developing a solution
- Testing and validating the model
Step 3: Interpretation and what-if analysis
- Analyzing the results and sensitivity analysis
- Implementing the results

Chapter 2: Linear Programming

Introduction to modelling - The Graphical Solution Method

Introduction
1) Define the decision variables
2) Formulate the objective function → minimize or maximize
3) Formulate the functional constraints
4) Formulate the sign restrictions → xi ≥ 0 or xi urs

Terminology
- Feasible solution
 Solution that satisfies all the problem’s constraints.
- Optimal solution
 Feasible solution that results in the largest possible objective function value when maximizing
(or smallest when minimizing).

Linear Programming Formulation

Three categories of constraints:
- Limited resources: “at most”
- Minimum performance: “at least”
- Conservation: “exactly”

Assumptions and Limitations of LP

- Certainty
➔ All parameters of the model are known constants.
- Linearity
➔ The objective function and constraints are expressed as linear functions.
1
 - Proportionality
➔ The contribution of each activity to the value of the objective function Z is proportional to the level
of the activity xj (cj ).
➔ The contribution of each activity to the left-hand side of each functional constraint is proportional
to the level of the activity xj (a ij ).
- Additivity
➔ Every function (in LP model) is the sum of the contributions of all individual decision variables.
➔ Can be violated due: complementary or competivity
- Divisibility
➔ Decision variables can have non-integer or fractional values (<-> integer programming)

The Graphical Solution Method

For linear programs with two or three variables.
Solution method
1) Determine the feasible region
- Draw the constraints one-by-one
- Find the set of points that satisfy all constraints
2) Determine optimal solutions
- Draw an objective function line
➢ Isoprofit line (maximization)
➢ Isocost line (minimization)
- Move the objective function line (draw parallel lines) → same slope
- Determine optimal solution
➢ Intersection objective function line and the feasible region
→ the feasible solution with the largest value is the optimal solution.

- Binding (equal) and nonbinding (unequal) constraints

- Feasible region: nonexistent, a single point, a line, a polygon or an unbounded area.
- Extreme points → optimal solution

Special cases
- Alternative or multiple optimal solutions
➔ In graphical method: objective function line is parallel to a boundary constraint.
- Infeasible LP
➔ LP is over-constrained → no feasible solutions → no optimal solution
- Unbounded LP
➔ Many feasible solutions → no optimal solution

The Simplex Method

The Principles
Optimality test
If a CFP solution (corner-point feasible solution) has no adjacent CPF solutions that are better, then it must be
an optimal solution.
➔ Choice of the next adjacent CPF solution is dependent on the rate of improvement in Z (largest rate
is chosen) that would be obtained by moving along the edge.

2
Setting Up the Simplex Method
Basic solution
- A basic solution is an augmented CPF solution (and includes slack variable values).
- Two basic solutions are adjacent if all but one of their non-basic variables are the same.
- Moving from one basic solution to another involves switching one variable from non-basic to basic and vice
versa for another variable.
➔ Entering variable: non-basic to basic
➔ Leaving variable: basic to non-basic

The Algebra of the Simplex Method

Rewriting the Linear Functions → Jordan-Gauss elimination: elementary row operations

The Simplex Method

Step 1:
If the problem is a minimization problem, multiply the objective function by -1.
Step 2:
If the problem formulation contains any constraints with negative RHS, multiply each constraint by -1.
Step 3:
Put the problem into standard form:
- Add a slack variable to each ≤ constraint.
- Subtract a surplus variable and add an artificial variable to each ≥ constraint.
- Set each slack and surplus variable’s coefficient in the objective function to zero.
Step 4:
Select the origin as initial basic solution:
- Select the decision variables to be the initial non-basic variables (set equal to zero).
Step 5:
Put the problem into canonical form:
- Add an artificial variable to each equality constraint and each ≥ constraint.
- Set each artificial variable’s coefficient in the objective function equal to -M, where M is a very large
number.
- Each slack and artificial variable becomes one of the basic variables in the initial basic solution.
- Conditions:
1) All right-hand sides (RHS) values are positive.
2) Each basic variable appears in 1 constraint.
3) Each constraint contains 1 basic variable.
4) The coefficients of the basic variables are all equal to 1.
Step 6:
Rewrite the objective function in terms of non-basic variables only such that the entering BV can be easily
determined. Hence, eliminate all BV’s from this row using elementary row operations.
Step 7:
Iterations:
1) Determine entering variable
 variable with the most negative value in the objective row
 corresponding column = pivot column
2) Determine leaving variable
 trade ratio = RHS / (exchange) coefficient
 select variable with smallest trade ratio
 corresponding row = pivot row
3
 3) Generate New Tableau
 Divide the pivot row by the pivot element
 Replace each non-pivot row
 Replace the objective row

Special cases
- Tie for the entering variable → choose arbitrarily
- Tie for the leaving variable → choose arbitrarily
- Degeneracy: basic variable with a value of zero = degenerated variable
- Alternative optimal solutions: non-basic variable with an objective row value equal to zero in the final
tableau → multiple optima
- Unboundedness: all entries in an entering column are non-positive → no leaving variable → entering
basic variable could be increased indefinitely → unbounded
- Infeasibility: there is an artificial variable in the optimal solution → problem is infeasible

Two-Phase Method
When using big M-method → split the problem and solve the problem in two phases

Phase 1: Divide the big M method objective function terms by M and drop the other negligible terms.
- Minimize Z = ∑ artificial variables
- subject to Revised constraints (with artificial variables)

Phase 2: Find the optimal solution for the real problem. Use the optimal solution of phase 1 as initial basic
feasible solution for applying the simplex method to the real problem (the big M method coefficients can be
dropped dependent of outcome of phase 1).
- Minimize Z = ∑ original variables
- subject to Original constraints (without artificial variables)

Duality Theory
Quick estimate of the objective function value → how to find the best possible upper bound?
Construct a linear combination of the constraints and multiply an unknown variable 𝑦𝑖 with each constraint.
Suppose that, instead of using our resources to produce products, we sell the available resources → what
prices should we charge?
➔ Decision variables 𝑦𝑖 : per unit charge for every resource.

Sensitivity Analysis
Sensitivity analysis, also referred to as post optimality analysis
Determine how the optimal solution in an LP is affected by changes, within specified ranges, in:
- The values of the objective function coefficients.
- The values of the right-hand side coefficients.
- The values of the exchange coefficients.

How might changes in the objective function coefficients affect the optimal solution?
The range of optimality of an objective function coefficient provides the range of values over which the current
solution remains optimal (keeping all other coefficients constant).

4
How might a change in the right-hand side for a constraint affect the feasible region and
perhaps cause a change in the optimal solution?
The range of feasibility of a right hand side coefficient is the range of that coefficient over which the shadow
price remains unchanged.

Shadow price vs Dual price:

The shadow price is the change in the objective function if the RHS is increased by 1.
The dual price gives the improvement in the objective function if the constraint is relaxed by one unit (RHS + 1
for ≤ constraints; RHS - 1 for ≥ constraints).
→ For ≤ constraints they are the same.
→ For ≥ constraints they have opposite signs.

Graphical Sensitivity Analysis

Graphically, the limits of a range of optimality are found by changing the slope of the objective function line
within the limits of the slopes of the binding constraint line:
𝑐1
- The slope of an objective function line: 𝑍 = 𝑐1 𝑥1 + 𝑐2 𝑥2 → −
𝑐2
𝑎1
- The slope of a constraint: 𝑎1 𝑥1 + 𝑎2 𝑥2 → −
𝑎2
Graphically the range of feasibility is determined by finding the values of a right hand side coefficient such that
the same two lines that determined the original optimal solution continue to determine the optimal solution
for the problem.

Sensitivity Analysis with Simplex Tableaus !!!

Change in a coefficient of a decision variable
- Decision variable is a non-basic variable
➔ Duality Theory
- Decision variable is a basic variable
➔ Fundamental Insight
- Objective function coefficient → range of optimality for 𝑐𝑘
- Technological coefficient → range of optimality for 𝑐𝑘
➔ When outside the range of optimality: simplex method
- Introduction of a new variable
➔ Consider new variable as a non-basic variable
➔ Duality Theory
- Introduction of a new constraint
➔ Check feasibility of current solution:
o If feasible: current solution remains optimal.
o If infeasible: add constraint to final simplex table and set slack variable as initial basic variable.

Change in a right-hand side coefficient

- Range of feasibility
➔ Fundamental Insight

Simultaneous changes in the parameters

100% rule
Calculate for each change the percentage of the allowable change (increase or decrease) for that coefficient. If
the sum of the percent changes in the coefficients does not exceed 100%, the original optimal solution will still
be optimal. (If the sum exceeds 100%, then we cannot be sure.)

5
Fundamental insight
For a problem with 2 constraints:
𝑅0𝐹𝐼𝑁𝐴𝐿 = 𝑅0𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑦1 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑦2 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿
𝑅1𝐹𝐼𝑁𝐴𝐿 = 𝑠1′ 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠2′ 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿
𝑅2𝐹𝐼𝑁𝐴𝐿 = 𝑠1′′ 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠2′′ 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿

For a problem with 3 constraints:

𝑅0𝐹𝐼𝑁𝐴𝐿 = 𝑅0𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑦1 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑦2 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑦3 𝑅3𝐼𝑁𝐼𝑇𝐼𝐴𝐿
𝑅1𝐹𝐼𝑁𝐴𝐿 = 𝑠1′ 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠2′ 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠3′ 𝑅3𝐼𝑁𝐼𝑇𝐼𝐴𝐿
𝑅2𝐹𝐼𝑁𝐴𝐿 = 𝑠1′′ 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠2′′ 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠3′′ 𝑅3𝐼𝑁𝐼𝑇𝐼𝐴𝐿
𝑅3𝐹𝐼𝑁𝐴𝐿 = 𝑠1′′′ 𝑅1𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠2′′′ 𝑅2𝐼𝑁𝐼𝑇𝐼𝐴𝐿 + 𝑠3′′′ 𝑅3𝐼𝑁𝐼𝑇𝐼𝐴𝐿

Extra example B
Question 4: Range of feasibility and sunk costs
- Given that aluminium is a sunk cost, what is the maximum amount the company should pay for 50 extra
pounds of aluminium?
- If aluminium were a relevant cost, what is the maximum amount the company should pay for 50 extra
pounds of aluminium?

Answer: Additional explanation of the theory

- When (resource) costs are sunk, the cost is not considered in the objective function.
- When (resource) costs are relevant, the cost is considered in the objective function.
- The shadow price reveals the increase in profit (= objective function value) if the resource availability (=
the right hand side of a constraint) is increased by one unit. The shadow price is thus the maximal price
that you want to pay (additionally) for an extra unit of the resource.
- If the cost is relevant, the cost is considered in the objective function, and hence, the shadow is the
additional amount that you want to pay for one unit of the resource. Thus, the price for one unit of the
resource = shadow price + variable cost (considered in the objective).
- If the cost is sunk, the cost is not considered in the objective function, and hence, the shadow price is
the amount that you want to pay for one unit of the resource. Thus the price for one unit of the
resource = shadow price + 0 (since this cost is not considered in the objective).
Answer:
- Since the cost for aluminium is a sunk cost, the shadow price provides the value of extra units of
aluminium. The shadow price for aluminium is the same as its dual price (for a maximisation problem).
The shadow price for aluminium is €3.125 per pound and the maximum allowable increase is 60 pounds.
Since 50 is in this range, then the €3.125 is valid, and the value of 50 additional pounds is €156.25.
- If the cost of aluminium were a relevant cost, the shadow price would be the amount above the normal
price of aluminium the company would be willing to pay. Thus if initially aluminium cost €4 per pound,
then additional units in the range of feasibility would be worth €4 + €3.125 = €7.125 per pound.

Multi-Criteria Decision Making

Goal Programming
Solve linear programs with multiple objectives and each objective viewed as a “goal”.
The goals themselves are added to the constraint set with 𝐝+ −
𝐢 and 𝐝𝐢 (deviation variables) acting as the
surplus and slack variables.
- Non-preemptive Goal Programming → all the goals are of roughly comparable importance
- Preemptive Goal Programming → hierarchy of priority levels for the goals

6
Formulate a goal programming model
1) Decide the priority of each goal
2) Decide the weight on each goal. If a priority level has more than one goal, for each goal i decide the
weight, 𝑤𝑖 , to be placed on the deviation(s), 𝑑𝑖+ and/or 𝑑𝑖− , from the goal.
3) Set up the initial linear program
➢ Min (𝑤1 𝑑1+ + 𝑤1 𝑑1− )
➢ s.t. Functional and Goal Constraints

The Transportation Problem

LP formulation special cases:
- Minimum shipping guarantee
- Maximum route capacity
- Unacceptable routes
Solution methods:
- General-purpose Simplex Method
- Transportation Simplex Method

The Transportation Simplex Method

Setting Up the Transportation Simplex Method
- The Transportation Simplex Tableau
➔ No artificial variables
➔ Dummy column if demand is less than supply
➔ Dummy row if supply is less than demand

Finding Basic Feasible Solutions for Transportation Problems

- The Northwest Corner Rule
 Step 1: Select the variable of the northwest corner of the transportation tableau (𝑥11 ).
- Minimum Cost Method
 Step 1: Select the variable with the smallest shipping cost.
- Vogel’s Approximation Method
 Step 1: For each row and column of the table remaining under consideration, find the difference
between the two lowest unit costs 𝑐𝑖𝑗 still remaining in that row or column (opportunity cost). In
that row or column having the largest opportunity cost (difference), select the variable with the
smallest remaining unit cost.

Obtaining the Optimal Solution using the Transportation Simplex Method

1) Optimality test
 A basic feasible solution is optimal if and only if the reduced cost ≥ 0 of every (i,j) such that 𝑥𝑖𝑗 is
non-basic.
2) Select an entering basic variable
- Calculate the reduced costs of the non-basic variables:
o MODI method: 𝑢𝑖 (row) and 𝑣𝑗 (column), set 𝑢1 = 0, remaining with 𝑐𝑖𝑗 = 𝑢𝑖 + 𝑣𝑗
→ improvement index: 𝐼𝑖𝑗 = 𝑐𝑖𝑗 − 𝑢𝑖 − 𝑣𝑗
o Stepping-Stone method: closed path (+ / –) → improvement index
→ EV = variable with largest negative reduced cost
- Select the variable with the largest negative reduced cost to enter the basis.
3) Determine the leaving basic variable
- Pivoting: closed path → + (recipient cells) and - (donor cells)

7
 → leaving variable = donor cell with smallest allocation
4) Determine the new basic feasible solution
 Add the value of the leaving basic variable to the allocation for each recipient cell and
subtracting this same amount from the allocation for each donor cell.

Sensitivity Analysis for Transportation Problems (exercise 6)

Changing the objective function coefficient 𝑐𝑖𝑗 of a non-basic variable (unoccupied cell) into 𝑐𝑖𝑗′
- The current basic solution remains feasible (RHS is not changed)
- The shadow prices 𝑢𝑖 and 𝑣𝑗 remain unchanged
 Range of values? Change 𝑐𝑖𝑗 to 𝑐𝑖𝑗 + ∆.
- Current basic solution remains optimal if 𝑐𝑖𝑗′ − 𝑢𝑖 − 𝑣𝑗 > 0

Changing the objective function coefficient 𝑐𝑖𝑗 of a basic variable (occupied cell) into 𝑐𝑖𝑗′
- The current basic solution remains feasible (RHS is not changed)
- Find new shadow prices 𝑢𝑖 and 𝑣𝑗 as the set of equations have changed
 Range of values? Change 𝑐𝑖𝑗 to 𝑐𝑖𝑗 + ∆ → than new shadow prices
- Current basic solution remains optimal as long as all non-basic variables price out nonnegative.

Increasing both supply 𝑠𝑖 and demand 𝑑𝑗 by ∆

- This change maintains a balanced transportation problem (Total supply = Total demand)
- If 𝑥𝑖𝑗 is a basic variable in the optimal solution then increase 𝑥𝑖𝑗 by ∆.
- If 𝑥𝑖𝑗 is a non-basic variable in the optimal solution then find the loop involving 𝑥𝑖𝑗 and some of the basic
variables. Mark this cell as the leaving variable and trace a closed path back to this non-basic variable,
alternately increasing and then decreasing the current basic variables in the loop by ∆.

The Assignment Problem

An assignment problem is a balanced transportation problem in which all supplies (assignees) and demands
(tasks) are equal to 1.
Objective:
Minimize the total cost assignment.
Methods:
- The Transportation Simplex method
- The general purpose Simplex Method
- The Network Simplex Method
Considerations:
- If the number of assignees does not equal the number of tasks, dummy assignees or tasks need to be
added with 0 assignment costs.
- If assignee i cannot perform task j, assign a large positive cost M to 𝑐𝑖𝑗 .
- To solve an assignment problem that maximizes the objective function, 2 methods:
1) Create an opportunity loss matrix:
→ Subtract all profits for each task from the maximum profit for that task and solve the
problem as a minimization problem.
2) Transform the profits matrix into a cost matrix
→ Multiply the profits through by -1 and solve the problem as a minimization problem.

The Hungarian Method

- Reduced cost matrix
- Solution method: see slides!
8
The Transshipment Problem
Transshipment problems are transportation problems in which a shipment may move through intermediate
nodes before reaching a particular destination node.
→ k transshipment points or junctions through which goods are shipped.

Chapter 3: Network Optimization Problems

The Shortest-Path Problem

Objective
Find the shortest path (the path with the minimum total distance) from the origin to the destination.
Dijkstra’s labeling algorithm
[𝑙𝑥 , 𝑘𝑥 ] → permanent label
(𝑙𝑥 , 𝑘𝑥 ) → tentative (= voorlopig) label.
With 𝑙𝑥 = distance from the source node, 𝑘𝑥 = preceding node

The Minimum Spanning Tree Problem

Objective
Choose the set of links that have the shortest total length among all sets of links that satisfy the condition that
a path between every pair of nodes is provided.
Tree: (n-1) links & n nodes
Prim’s algorithm
Cheapest arc between a marked node and an unmarked node.
Kruskal’s algorithm
Cheapest arc between:
- Two unmarked nodes.
- A marked node and an unmarked node.
- Two marked nodes that are not connected.

The Maximum Flow Problem

Objective
Maximize the total amount of flow from the source to the sink.
Caution!
In case of multiple sources and/or multiple sinks, create a single dummy source out of which all flow originates
and/or dummy sink that absorbs all the flow that terminates at the sinks.
Ford-Fulkerson algorithm
Step 1: Find any directed path from the source to the sink that has a strictly positive flow capacity remaining (=
identify an augmenting path). Stop if no such path exists, the optimal solution is found.
Step 2: Determine f, the maximum flow along this path, given by the smallest flow capacity of all arcs in the
path. Increase the flow through the network by sending the amount f over this path.
Step 3: Subtract f from the remaining flow capacity in the forward direction for each arc in the path, and add f
to the remaining flow capacity in the backward direction for each arc in the path. Go to Step 1.

The Minimum Cut Problem

Objective
Find the cut that has the minimum cut value.

9
Solution method
Ford-Fulkerson algorithm
Max flow min cut theorem
For any network having a single origin node and a single destination node, the maximum possible flow from
origin to destination equals the minimum cut value for all cuts in the network.

Weak duality property

Maximum flow value ≤ capacity of any cut
→ capacity of any cut is an upper bound on the maximum flow.
Strong Duality Property
Maximum flow value = Capacity of minimum cut

The Minimum Cost Flow Problem

Objective
Minimize the total cost of sending the available supply through the network to satisfy the demand.

Complex Network Optimization Problems

Travelling Salesman Problem (TSP)
Make a tour at minimal costs that visits all customers exactly once.

Intermezzo
BIP Encoding
Use Big M to turn-off constraint:
Either:
- 3𝑥1 + 2𝑥2 < 18
- and 𝑥1 + 4𝑥2 < 16 + M (and M is very BIG)
Or:
- 3𝑥1 + 2𝑥2 < 18 + M
- and 𝑥1 + 4𝑥2 < 16

Use binary y to decide which constraint to turn off:

- 3𝑥1 + 2𝑥2 < 18 + y M
- 𝑥1 + 2𝑥2 < 16 + (1-y)M
- y ∈ {0,1}

LP Encoding
Introduce 𝑦𝑖 to turn off each constraint i

Use Big M to turn-off constraint:

- 𝑓1 (𝑥1 , ... , 𝑥𝑛 ) < 𝑑1 + M𝑦1 ,
- 𝑓𝑁 (𝑥1 , … , 𝑥𝑛 ) < 𝑑𝑁 + M𝑦𝑁

Constrain K of the 𝑦𝑖 to select constraints:

- K out of N constraints hold: ∑𝑁 𝑖 = 1 𝑦𝑖 = N - K
𝑁
- At least K of N hold: ∑𝑖 = 1 𝑦𝑖 ≤ N – K

10
Chapter 4: Integer Programming

Integer Programming
http://web.mit.edu/15.053/www/AMP-Chapter-09.pdf

Integrality requirement
One or more variables must assume an integer value.

Types of integer linear programming models

- Pure integer (linear) program (PIP) → All integer LP
➢ A linear program in which all variables are required to be integers.
- Binary integer program (BIP)
➢ All variables in an LP are restricted to be 0 or 1 → binary variables
- Mixed integer (linear) program (MIP)
➢ Only a subset of the variables are restricted to be integer or binary.

Applicability of Integer Programming

The use of discrete variables
- Indivisible (discrete) quantities
- Yes-or-no (binary) decision variables
- Indicator variables 𝒚𝒊 : binary variables that are linked to continuous variables to indicate certain states.

IP Formulation
Formulating logical constraints: see slides 25-33
- Contingent decisions: A implies B
- Pack constraint: at most one out of n items can be selected.
- Partition constraint: exactly one out of n items must be selected.
- Cover constraint: at least one out of n items must be selected.
- Cardinality constraint: more than k out of n items must be selected.
- K out of the N constraints must hold → indicator variables 𝑦𝑖 whether or not constraint i is satisfied.
- Variable upper bound constraint: “If you build a warehouse, you can store up to 13 tons of x in it.”
- Semi-continuous variable: “If you build a warehouse, it must be used to store at least 56 tons but no
more than 141 tons.”
- Binary factorization
- The knapsack problem
o n items, with per unit values and weights 𝑣𝑖 and 𝑤𝑖 (i = 1, …, n), can be put into a bag that can carry a
maximal weight of W.
o How many of each item should be put into the bag such that the total value is maximized?
- The fixed-charge problem → a fixed or lump-sum cost
- Blending problem

Solution methods
Rounding the LP solution
LP relaxation (LPR):
- Ignoring the integrality constraints.
- LP solution
- Rounding up or down to integer values → may be infeasible or sub-optimal
➔ Complete enumeration of feasible ILP solutions

11
Enumeration procedure
Make sure that only a small fraction of the feasible solutions need to be examined:
- Dynamic programming (see later)
- Branch-and-bound: divide-and-conquer method
➢ Initialisation: set 𝑍 ∗ = −∞
➢ Stopping rules:
o Optimality test: stop when there are no remaining subproblems to investigate, the current
incumbent is optimal.
o Specify a sub-optimality tolerance factor: allows to stop once an integer solution is found
that is within some % of the optimal value.
➢ Branch-and-bound with binary variables:
o Dividing (branching): partition the set of feasible solutions into smaller subsets.
▪ Branching: partition into subsets by focusing on one variable.
o Conquering (bounding + fathoming): find a bound (Z*) for how good the best solution in a
subset can be and then discard the subset if its bound indicates that it can’t contain an
optimal solution.
▪ Bounding: a bound for sub-problems is obtained by solving the LP relaxation. (Note that
the optimum of an IP can never be better than the optimum of its LP relaxation.)
▪ Fathoming: dismissing a subproblem from further consideration.
• The LP relaxed solution is worse than the best known integer solution.
• The LP relaxation is infeasible.
• The LP relaxed solution is all integer (update best known integer solution).
➢ Branch-and-bound with general integer variables:
o Branching: select the first integer-restricted variable with a fractional value in the LP relaxed
solution and create two new subproblems.
o Bounding: calculate the optimum of the LP relaxation.
o Fathoming: dismiss a subproblem if
▪ The LP solution is worse than the best known integer solution.
▪ The LP relaxation is infeasible.
▪ The LP solution is integer.
Branch-and-cut for BIP
- Automatic BIP preprocessing
➢ Fixing variables
o If one value of a variable cannot satisfy a constraint (even when the other variables equal their best value for trying
to satisfy the constraint), then that variable can be fixed at its other value.
➢ Eliminating redundant constraints
o If a functional constraint satisfies even the most challenging binary solution, then it has been made redundant by
the binary constraints and can be eliminated.
➢ Tightening ≤ constraints
o A constraint can be tightened by adjusting the coefficients and right-hand side such that the feasible region of the LP
relaxation reduces without eliminating feasible binary solutions.
- Generation of cutting planes
➢ A cutting plane (or cut or valid inequality) = a new functional constraint that reduces the feasible
region for the LP relaxation without eliminating feasible solutions for the IP problem.
➢ Minimum cover cuts (for BIP) → for functional constraints in ≤ form with only nonnegative
coefficients.
➢ Gomory cuts (for PIP) → when all coefficients and right-hand sides have integer values.
o Gomory cuts make the optimum of the LP relaxation infeasible by adding a constraint that
cuts the LP optimum from the solution space, based on the fractional values that appear in
the LP optimum.

12
Chapter 5: Dynamic Programming

General Description
General description
The problem can be divided in N separate stages with a policy decision required at each stage.
Each stage has a number of states associated with the beginning of that stage.
- In stage n the system has a certain state 𝑥𝑛 .
The effect of the policy decision 𝒅𝒏 at stage n is to transform the current state 𝑥𝑛 to a state 𝑥𝑛+1 associated
with the beginning of the next stage.
- The return function 𝑟𝑛 (𝑥𝑛 , 𝑑𝑛 ) gives the ‘return’ of making decision 𝑑𝑛 being in state 𝑥𝑛 at stage n.
- The state transition function 𝑡𝑛 (𝑥𝑛 , 𝑑𝑛 ) gives the state 𝑥𝑛+1 at stage n+1 that is reached after making
decision 𝑑𝑛 in state 𝑥𝑛 at stage n.

Recursive relationship
The cumulative return is the sum of the immediate return in the current stage and the cumulative return in the
remaining stages:
∗
- 𝑓𝑛 (𝑥𝑛 , 𝑑𝑛 ) = 𝑟𝑛 (𝑥𝑛 ,𝑑𝑛 ) + 𝑓𝑛+1 (𝑥𝑛+1 ) with 𝑥𝑛+1 = 𝑡𝑛 (𝑥𝑛 ,𝑑𝑛 )
The optimal cumulative return gives the best cumulative return in each state:
- 𝑓𝑛∗ (𝑥𝑛 ) = 𝑚𝑎𝑥/𝑚𝑖𝑛𝑑𝑛 (𝑓𝑛 (𝑥𝑛 ,𝑑𝑛 ))
In the final stage N, there is no next stage, so the cumulative return is equal to the immediate return:
- 𝑓𝑁 (𝑥𝑁 , 𝑑𝑁 ) = 𝑟𝑁 (𝑥𝑁 , 𝑑𝑁 )
- 𝑓𝑁∗ (𝑥𝑁 ) = 𝑚𝑎𝑥/𝑚𝑖𝑛𝑑𝑁 (𝑟𝑁 (𝑥𝑁 , 𝑑𝑁 ))
Backward recursion: to solve the problem, start in stage N and work back to stage 1.

Table for exercises:

Chapter 6: Non Linear Programming

The general form of a nonlinear programming problem is to find x = (𝑥1 , 𝑥2 , …, 𝑥𝑛 ) so as to:
- Maximize f(x)
- subject to: 𝑔 𝑖 (x) ≤ 𝑏 𝑖 for i = 1, 2, …, m
x ≥0
where f(x) and the 𝑔 𝑖 (x) are given functions of the n decision variables.

One-variable unconstrained optimization

The solution to the one-variable unconstrained NLP (maximize {f(x): x∈[a, b]}) may occur at one of the four
following cases:
- Case 1: df(x) exists and reaches a value of 0 within the range [a, b].
- ANALTYICAL procedures:
- For a concave function f(x), x = x* is a global maximum if and only if df(x)/dx = 0 at x = x*.
- For a convex function f(x), x = x* is a global minimum if and only if df(x)/dx = 0 at x = x*.
- Case 2: df(x) does not exist (i.e. at a discontinuity).
 Test points on either side of the discontinuity.
- Case 3: the endpoints a and b of the interval [a, b].
- Case 4: for nonlinear functions f(x), the derivative df(x)/dx may still be nonlinear and you may not be able
to solve the equation analytically.
13
 - NUMERICAL search procedures: see slides 29-34
𝑥+𝑥
- Bisection method → 𝑥 ′ = → STOP when 𝑥̅ − 𝑥 ≤ 2ε
2
Table:

′
𝑓 (𝑥𝑖)
- Newton’s method → 𝑥𝑖+1 = 𝑥𝑖 − ′′ → STOP when |𝑥𝑖+1 – 𝑥𝑖 | ≤ ε
𝑓 (𝑥𝑖)
Table:

Multivariable unconstrained optimization

𝜕𝑓(𝑥 ∗ )
The point x* = (𝑥1∗ ,𝑥2∗ , … ,𝑥𝑛∗ ) is called a stationary point if it satisfies: = 0 for all i = 1, …, m.
𝜕𝑥𝑖
Gradient 𝛁𝒇(𝒙) = vector of all partial derivatives → defines direction in which rate increase of f(x) is maximal.

Gradient search
Step 1: Initialization
- Select ε and any trial solution 𝑥0 .
Step 2:
- If all partial derivatives in the solution are smaller than ε, stop.
Step 3: We want to move as far as possible, say t* units, in this direction.
- Thus, select the next point 𝒙𝒎+𝟏 = 𝒙𝒎 + t*∇f(𝒙𝒎 ).
- Moving in the direction of the gradient ∇f(𝑥𝑚 ) increases the function value f.
- Therefore, express f(𝒙𝒎 + t ∇ f(𝒙𝒎 )) as a function of t.
Step 4: How big can t* be?
- Use one-variable unconstrained optimization to find t = t* that maximizes f(𝒙𝒎 + t ∇ f(𝒙𝒎 )) over t > 0.
- Note: If it is a minimization problem, we minimize f(𝒙𝒎 + t*∇f(𝒙𝒎 )) subject to the restriction that t* > 0.
Step 5:
- Update the solution: x’ ≡ x’ + t*∇f(x’).
Step 6:
- Once we have found our new point 𝑥𝑚+1 = 𝑥𝑚 + t*∇f(𝑥𝑚 ), we evaluate ||∇f(𝑥𝑚+1 )||. If the length of the
gradient is less than 0.01, then we have found a local maximum and can stop. Otherwise, return to step 2.
Table:

Constrained optimization
Consider the following NLP problem type, in which the restrictions are equality constraints (CNLP = ):
- Maximize (Minimize) z = f(x1 , x2 , …, xn )
- Subject: g i (x1 ,x2 , …,xn ) = bi , i = 1,…,m

Lagrangian relaxation
= convert the constrained NLP= into an unconstrained NLP and solve it as usual.
Lagrangian function:
𝐿(𝑥, 𝜆) = 𝑓(𝑥) + ∑𝑚
𝑖=1 𝜆𝑖 [(𝑏𝑖 − 𝑔𝑖 (𝑥)] with 𝜆𝑖 = Lagrangian multiplier for constraint i.
➔ Lagrangian multiplier 𝛌𝐢 : rate of change in the optimal value of f(x) per unit increase in bi .
14
Karush-Kuhn-Tucker conditions
x = (x1 , x2 , …, xn ) can be an optimal solution for the nonlinear programming problem (NLP)
Maximize f(x)
s.t. 𝑔𝑖 (𝑥) ≤ 𝑏𝑖 , for all i = 1, 2, …, m
𝑥 ≥ 0.
if there exist m numbers u1 , u2 , …, u𝑚 such that all the KKT conditions are satisfied.
If f(x) is concave and all g i (x) are convex, then a solution satisfying the KKT conditions is optimal.
𝑑𝑓′(𝑥,𝜆)
1) 𝑑𝑥
≤0 ∀𝑥
𝑑𝑓′(𝑥,𝜆)
2) 𝑥 𝑑𝑥
= 0 ∀𝑥

3) 𝑔(𝑥) ≤ 𝑏 → 𝑔(𝑥) − 𝑏 ≤ 0
4) 𝜆 (𝑔(𝑥) − 𝑏) = 0
5) 𝑥 ≥ 0
6) 𝜆 ≥ 0

Chapter 7: Decision Making Under Uncertainty

Decision Analysis
Decision making without experimentation
Decision making under uncertainty
- Maximax (minimin) criterion (optimistic approach)
o Choose the alternative with the largest maximum payoff.
o If the payoff table was in terms of costs, the decision with the lowest cost would be chosen.
- Maximin (minimax) criterion (conservative approach)
o For each decision the minimum payoff is listed and then the decision corresponding to the maximum
of these minimum payoffs is selected (choose the alternative with the largest minimum payoff).
o If the payoff was in terms of costs, the maximum costs would be determined for each decision and
then the decision corresponding to the minimum of these maximum costs is selected.
- Equally likely criterion (or Laplace criterion)
o Choose the alternative with the largest average payoff. The average payoff for every alternative is
calculated and the one with maximum average payoff is picked.
o This criterion assumes that probabilities of occurrence for the states of nature are equal.
- Criterion of realism (Hurwicz’ criterion or weighted average)
o Select coefficient of realism α between 0 and 1, then choose the alternative with the largest value of
𝛼(𝑚𝑎𝑥𝑖𝑚𝑎𝑙 𝑝𝑎𝑦𝑜𝑓𝑓 𝑓𝑜𝑟 𝑡ℎ𝑒 𝑎𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒)
+ (1 − 𝛼)(𝑚𝑖𝑛𝑖𝑚𝑎𝑙 𝑝𝑎𝑦𝑜𝑓𝑓 𝑓𝑜𝑟 𝑡ℎ𝑒 𝑎𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒)
o When α is close to 1, the decision maker is more optimistic about future; and when α is close to 0,
the decision maker is more pessimistic about future.
- Minimax regret criterion (or Savage regret criterion)
o This criterion requires construction of the opportunity loss table which is done by calculating for each
state of nature the difference between each payoff and the largest payoff for that state of nature.
o The maximum regret for each possible decision is listed.
o Choose the alternative with the smallest maximum opportunity loss (= the minimum of the
maximum regrets).

15
Decision making under risk → prior probabilities
- Maximum likelihood criterion
o Identify the most likely state of nature and choose the alternative with the largest payoff for this
state of nature.
- Bayes’ rule (Expected Value approach)
o The expected return for each decision is calculated by summing the products of the payoff under
each state of nature and the probability of the respective state of nature occurring.
o Choose the alternative with the maximum expected payoff or expected monetary value (EMV):

𝐸𝑀𝑉𝑖 = ∑ 𝑟𝑖𝑗 𝑝𝑗
𝑗

with 𝑝𝑗 probability of the 𝑗𝑡ℎ state of nature,

𝑟𝑖𝑗 payoff of 𝑖 𝑡ℎ alternative under the 𝑗𝑡ℎ state of nature.
- Expected regret approach
o Choose the alternative with the smallest expected regret or opportunity loss (EOL):

𝐸𝑂𝐿𝑖 = ∑ 𝑔𝑖𝑗 𝑝𝑗
𝑗

with 𝑝𝑗 probability of the 𝑗𝑡ℎ state of nature,

𝑔𝑖𝑗 regret for the 𝑖 𝑡ℎ alternative under the 𝑗𝑡ℎ state of nature.
- The decision with the largest EMV will also have the smallest EOL.

Expected value of perfect information

The expected value of perfect information (EVPI) is the increase in expected payoff that would result from
knowing with certainty which state of nature will occur.
→ upper bound on the expected value of any sampling or survey information.

EVPI calculation:
1) Determine the optimal return (with PI) corresponding to each state of nature.
2) Compute the expected value of these optimal returns.
3) Subtract the EV of the optimal decision from this expected value (determined in step 2):
EVPI = EV with PI - maximum EV

Decision making with experimentation (sample information)

Bayes’ theorem
Given:
- n possible states of nature (indexed by i)
- m possible test findings (indexed by j)
- Prior probabilities P(state = i), for all i.
- Conditional probabilities P(finding = j | state = i), for all i and j.
Outcome:
- Posterior probabilities P(state = i | finding = j), for all i and j.
Calculate:
𝑃(𝑆𝑖 ∩ 𝐹𝑗 ) 𝑃(𝐹𝑗 | 𝑆𝑖 )𝑃(𝑆𝑖 )
𝑃(𝑆𝑖 | 𝐹𝑗 ) = =
𝑃(𝐹𝑗 ) ∑𝑖 𝑃(𝐹𝑗 | 𝑆𝑖 )𝑃(𝑆𝑖 )

16
Posterior probabilities calculation:
1) For each state i, multiply its prior probability by the conditional probability for the finding j (indicator).
This gives the joint probabilities for states and findings.
2) Sum these joint probabilities over all states. This gives the marginal probability for each finding.
3) For each state i, divide its joint probability with finding j by the marginal probability for that finding.
This gives the posterior probability.

Expected value of experimentation (sample information)

The additional expected profit possible through knowledge of the sample or survey information.
EVSI Calculation:
1) Determine the optimal decision and its expected return for the possible outcomes of the sample or
survey using the posterior probabilities for the states of nature.
2) Compute the expected value of these optimal returns.
3) Subtract the EV of the optimal decision obtained without using the sample information from the
amount determined in Step 2.
EVSI = EV with experimentation – EV without experimentation

Decision trees
A decision tree has two types of nodes:
- The decision nodes (square nodes) correspond to decision alternatives: a decision must be made.
- The event nodes (round nodes) correspond to the states of nature: a random event will occur.
A decision tree has two types of branches:
- Branches leaving decision nodes represent the different decision alternatives, from which one must be
selected.
- Branches leaving event nodes represent the different states of nature, each occurring with a given
probability.

Markov Chains
Introduction: Stochastic processes
A stochastic process is a series of random variables {𝑋𝑡 }, with index t running through a given set T. The values
𝑋𝑡 can assume are called the states of the process. We assume that the number of possible states M is finite.
If the next state 𝑋𝑡+1 of the process depends only on the present state 𝑋𝑡 and is independent of past history,
the process is called a Markov Chain.
The process will move from a state to other states with known transition probabilities.

Markov chains
The evolution over time of a (discrete time) stochastic process is described by the (one-step) transition
probabilities:
- P(𝑋𝑡+1 = j | 𝑋𝑡 = i, 𝑋𝑡−1 = 𝑠𝑡−1 , ..., 𝑋0 = 𝑠0 )
A stochastic process has the Markovian property if the transition probabilities depend on the current state
only:
- P(𝑋𝑡+1 = j | 𝑋𝑡 = i, 𝑋𝑡−1 = 𝑠𝑡−1 , ..., 𝑋0 = 𝑠0 ) = P(𝑋𝑡+1 = j | 𝑋𝑡 = i)
A stochastic process is a Markov chain if it has the Markovian property.
If the transition probabilities do not change over time, they are stationary:
- P(𝑋𝑡+1 = j | 𝑋𝑡 = i) = P(𝑋1 = j | 𝑋0 = i) = 𝑝𝑖𝑗
This implies that the n-step transition probabilities are also stationary:
(𝑛)
- P(𝑋𝑡+𝑛 = j | 𝑋𝑡 = i) = P(𝑋𝑛 = j | 𝑋0 = i) = 𝑝𝑖𝑗

17
Chapman-Kolmogorov equations
The n-step probabilities can be obtained recursively from the one-step transition probabilities (with P the
transition matrix).
The n-step transition probability matrix P(n) is given by the n’th power of P: 𝑃(𝑛) = 𝑃(𝑚−𝑘) × 𝑃(𝑘)
Steady-State Probability 𝝅𝒋 for each state j is the long-run probability that a process starting in any state will
be found in state j.

Classification of states in a Markov chain

States i and j are said to communicate if they are accessible from one another.
A Markov chain is irreducible if all states communicate, i.e. if there is only one class.
Transient states
A state is transient if the process may never return to it after visiting it. In other words, state i is transient if
there exists a state j that is accessible from i, but not vice versa.
Recurrent states
A state is recurrent if the process will certainly return to it after visiting it. So, a state is recurrent if and only if it
is not transient.
A state i is absorbing if the process will never leave to another state once it enters state i. An absorbing state is
a special case of a recurrent state (𝑝𝑖𝑖 = 1).
Periodicity
(𝑛)
The period of state i is the integer t (t > 1) such that 𝑝𝑖𝑖 = 0 for all values of n other than t, 2t, 3t, …
If there are two consecutive numbers s and s+1 such that the process can be in state i at times s and s+1, the
state is called aperiodic.
Aperiodic, recurrent states are called ergodic states.

Long-run properties of Markov chains

The steady-state probabilities 𝝅𝒋 can be calculated from 𝝅 = 𝝅 x P and 𝝅𝟏 +𝝅𝟐 + ... + 𝝅𝑴 = 1, i.e. the steady-
state probabilities do not change from transition to transition and they cumulate to 1.

𝜋0 = 𝜋0 𝑃00 + 𝜋1 𝑃10
𝜋1 = 𝜋0 𝑃01 + 𝜋1 𝑃11
𝜋0 + 𝜋1 = 1

Expected average cost

The (long-run) expected average cost per unit time:
n M
1
lim E [ ∑ C(Xt) ] = ∑ πj C(j)
n→∞ n
t=1 j=0

Other Markov Chain Concepts

The first passage time is the number of transitions (number of steps) necessary in going from state i to j for the
first time.
➔ Expected first passage time 𝜇𝑖𝑗 :
𝜇𝑖𝑗 = 1 + ∑ 𝑝𝑖𝑘 𝜇𝑘𝑗
𝑘≠𝑗

The recurrence time is the number of transitions (number of steps) necessary to return to state i when starting
from state i.
➔ Expected recurrence time 𝜇𝑖𝑖 :
1
𝜇𝑖𝑖 =
𝜋𝑖
18
Absorbing states
If there is more than one absorbing state in a Markov chain, the probability of going to absorbing state k from
initial state i is the probability of absorption fik .
Absorption probabilities 𝐟𝐢𝐤 for state k:
𝑀

𝑓𝑖𝑘 = ∑ 𝑝𝑖𝑗 𝑓𝑗𝑘

𝑗=0

𝑤𝑖𝑡ℎ 𝑓𝑘𝑘 = 1 𝑎𝑛𝑑 𝑓𝑖𝑘 = 0 𝑖𝑓 𝑠𝑡𝑎𝑡𝑒 𝑖 𝑖𝑠 𝑟𝑒𝑐𝑢𝑟𝑟𝑒𝑛𝑡 𝑎𝑛𝑑 𝑖 ≠ 𝑘.

Transition Matrix with Submatrices

If a Markov chain has both absorbing and non-absorbing states, the states may be rearranged so that the
transition matrix can be written as the following composition of four submatrices: I, 0, R, and Q:
- I = an identity matrix indicating one always remains in an absorbing state once it is reached.
- 0 = a zero matrix representing 0 probability of transitioning from the absorbing states to the non-
absorbing states.
- R = the transition probabilities from the non-absorbing states to the absorbing states.
- Q = the transition probabilities between the non-absorbing states.
The fundamental matrix, N, is the inverse of the difference between the identity matrix and the Q matrix:
𝑁 = (𝐼 − 𝑄)−1
The 𝑵 · 𝑹 matrix, the product of the fundamental matrix and the R matrix, gives the probabilities of eventually
moving from each non-absorbing state to each absorbing state.

Main questions on Exam!

- LP analysis (sensitivity analysis)
- Network optimization problems (Transportation, Transshipment, Assignment, Shortest path, Minimum
spanning tree, etc.)
- IP modelling

19