INDU 342: Logistics Networks Models

Design of a logistics system

Claudio Contardo

Mechanical, Industrial and Aerospace Engineering

Concordia University

Lecture 6
Discrete facility location problems

X X
Minimize Z = fi yi + cij sij
i∈V1 i∈V1 ,j∈V2

subject to
X
sij ≤ qi yi i ∈ V1
j∈V2
X
sij = dj j ∈ V2
i∈V1

s≥0
y binary.
Remarkable cases of discrete facility location problems
Uncapacitated FLP
Non capacitated facilities can be given a dummy capacity
X
qi = Q = dj
j∈V2

X X
Minimize Z = fi yi + cij sij
i∈V1 i∈V1 ,j∈V2

subject to
X
sij ≤ Qyi i ∈ V1
j∈V2
X
sij = dj j ∈ V2
i∈V1

s≥0
y binary.
Remarkable cases of discrete facility location problems

Integrality property of the Uncapacitated FLP

The UFLP admits an optimal solution (y ∗ , s∗ ) such that sij ∈ {0, dj } for
every i ∈ V1 , j ∈ V2

Non integrality of the Capacitated FLP

When capacities (qi )i∈V1 are such that qi < Q the above statement is no
longer true
Remarkable cases of discrete facility location problems

Multiple-source facility location problem

Using capacities qi < Q leads to the Capacitated FLP (CFLP) where it is
allowed for the flows sij to be split among multiple facilities

X X
Minimize Z = fi yi + cij sij
i∈V1 i∈V1 ,j∈V2

subject to
X
sij ≤ qi yi i ∈ V1
j∈V2
X
sij = dj j ∈ V2
i∈V1

s≥0
y binary.
Remarkable cases of discrete facility location problems
Single-source facility location problem
When in addition of capacities, it is required that every node be allocated
to an unique facility (which must then meet ts entire demand), we refer
to the Single-source capacitated FLP (SSCFLP)

X X
Minimize Z= fi yi + cij sij
i∈V1 i∈V1 ,j∈V2

subject to
X
sij ≤ qi yi i ∈ V1
j∈V2
X
sij = dj j ∈ V2
i∈V1

sij ∈ {0, dj } i ∈ V1 , j ∈ V2
y binary.
Remarkable cases of discrete facility location problems

Single-source facility location problem

It is convenient to define xij = sij /dj , γij = cij dj and then write the
problem as below

X X
Minimize Z = fi yi + γij xij
i∈V1 i∈V1 ,j∈V2

subject to
X
dj xij ≤ qi yi i ∈ V1
j∈V2
X
xij = 1 j ∈ V2
i∈V1

xij ∈ {0, 1} i ∈ V1 , j ∈ V2
y binary.
A Lagrangean heuristic for the CFLP

Relax the P
demand satisfaction constraints. This is, instead of
imposing P i xij = 1, allow for them to be violated and let
vj = 1 − i xij be the level of violation (> 0 meaning demand is
unmet, < 0 meaning excess of demand is met)
Delete the Passociated constraints and extend the objective as
Z ′ = Z + j λj vj for a certain λ ∈ R|V2 | (the Lagrangean
multipliers)
We can
P in addition delete the variables vj and replace them by
1 − i xij under the condition that 0 ≤ xij ≤ 1 for every i, j
A Lagrangean heuristic for the CFLP

The associated relaxed problem becomes

!
X X X X
Minimize Z(λ) = fi yi + γij xij + λj 1− xij
i∈V1 i∈V1 ,j∈V2 j∈V2 i∈V1

subject to
X
dj xij ≤ qi yi i ∈ V1
j∈V2

0 ≤ xij ≤ 1 i ∈ V1 , j ∈ V2
y binary.
A Lagrangean heuristic for the CFLP

The objective can be rewritten as

!
X X X X
Minimize Z(λ) = fi yi + γij xij + λj 1− xij
i∈V1 i∈V1 ,j∈V2 j∈V2 i∈V1
 
X X X
= λj + fi yi + (γij − λj ) xij 
j∈V2 i∈V1 j∈V2

Separability
The Lagrangean problem can now be separated by potential facility
A Lagrangean heuristic for the CFLP

For a given set of multipliers γ and potential facility i ∈ V1 we solve the

problem
X
Minimize Z(λ, i) = fi yi + (γij − λj ) xij
xi· ,yi
j∈V2

subject to
X
dj xij ≤ qi yi
j∈V2

0 ≤ xij ≤ 1 j ∈ V2
yi binary.
A Lagrangean heuristic for the CFLP

To solve this problem (and compute Z(λ, i)) we proceed as follows:

For yi = 0, we have that xij = 0 for every j ∈ V2 and then the
optimal value of the MILP is zero
For yi = 1, the problem resorts to a continuous knapsack problem
that can be solved via a sorting algorithm (very quick!)
We select the value that leads to the minimum objective
A Lagrangean heuristic for the CFLP

Finding a subset of feasible facilities

Sort the facilities by non-decreasing value of Z(λ, i)
Pl P
Let k = min{l : t=1 qil ≥ j dj }
Select I = {i1 , . . . , ik } as the open facilities
Find an optimal assignment of the customers in V2 to the facilities
in S by solving a LP
A Lagrangean heuristic for the CFLP

The optimal assignment of the customers in V2 to the facilities in S is

given by
X
Minimize ZASS (S) = γij xij
x
i∈S,j∈V2

subject to
X
dj xij ≤ qi i∈S
j∈V2

0 ≤ xij ≤ 1 i ∈ S, j ∈ V2 .
A Lagrangean heuristic for the CFLP

Let x∗ij be the solutions identified by the different subproblems Z(λ, i),
and let α > 0
Identify all the j ∈ V2 such that
X
vj = 1 − x∗ij ̸= 0
i∈V1

If vj > 0, increase the value of λj by making λj ← λj + αvj

Else, if vj < 0, decrease the value of λj by making λj ← λj + αvj
Solve the problem again with the updated Lagrangean multipliers
For the process to converge, the step α needs to be reduced in every
iteration (actually, tend to zero) but the sum must diverge to +∞
1
For example, consider α(k) = k
Single-commodity two-echelon location

Suppose now that flows are transported in a three-layer network, as

follows
Nodes
Three sets of nodes:
Source nodes V1
Potential facilities V2
Final customers V3

Flows
Source nodes shipping flows to intermediate facilities
Potential intermediate facilities shipping flows to final customers
Single-commodity two-echelon location

Quantities
Each source node i ∈ V1 has a certain supply oi
Each intermediate node j ∈ V2 has a certain capacity qj
Each final customer node k ∈ V3 has a certain demand dk

Costs
Unit transportation cost cijk , i ∈ V1 , j ∈ V2 , k ∈ V3
Fixed location costs fj , j ∈ V2
Single-commodity two-echelon location

Decision variables
Location variables yj ∈ {0, 1} for every j ∈ V2
Transportation variables sijk ≥ 0 to indicate the amount of flow
being sent from i to k through the intermediate facility j
Single-commodity two-echelon location
We can formulate this problem as follows:
X X
Minimize Z = fj yj + cijk sijk
s,y
j∈V2 i∈V1 ,j∈V2 ,k∈V3

subject to
X
sijk ≤ oi i ∈ V1
j∈V2 ,k∈V3
X
sijk ≤ qj yj j ∈ V2
i∈V1 ,k∈V3
X
sijk = dk k ∈ V3
i∈V1 ,j∈V2

s≥0
y binary.
Multi-commodity two-echelon location

Let us consider the two-echelon location problem but now multiple

commodities (h ∈ H) are to be transported from sources (V1 ) to final
destinations (V3 ) via intermediate facilities (V2 )

Quantities
Each source node i ∈ V1 has a certain supply oih of product h ∈ H
Each final customer node k ∈ V3 has a certain demand dkh of
product h

Costs
Unit transportation cost cijkh , i ∈ V1 , j ∈ V2 , k ∈ V3 , h ∈ H
Multi-commodity two-echelon location

Decision variables
Location variables yj ∈ {0, 1} for every j ∈ V2
Transportation variables sijkh ≥ 0 to indicate the amount of flow of
product h being sent from i to k through the intermediate facility j
Multi-commodity two-echelon location
We can formulate this problem as follows:
X X
Minimize Z = fj yj + cijkh sijkh
s,y
j∈V2 i∈V1 ,j∈V2 ,k∈V3 ,h∈H

subject to
X
sijkh ≤ oih i ∈ V1 , h ∈ H
j∈V2 ,k∈V3
X
sijkh ≤ qj yj j ∈ V2
i∈V1 ,k∈V3 ,h∈H
X
sijkh = dkh k ∈ V3 , h ∈ H
i∈V1 ,j∈V2

s≥0
y binary.
Location-covering problems

We are given
A set of potential locations V1 , fixed costs fi , i ∈ V1
A set of customer locations V2
A 0-1 incidence matrix A such that for every i ∈ V1 , j ∈ V2 , aij = 1
indicates whether location i covers the customer j

Remarkable examples
Find the optimal location of public primary schools so every postal
code is away of at most 500m from the closest school
Locate fire stations so every postal code is at most 2 km away from
the closest fire station
Location-covering problems

Variables
yi ∈ {0, 1} indicating whether a facility is open at location i

We can formulate this problem as follows:

X
Minimize Z = fi yi
y
i∈V1

subject to
X
aij yj ≥ 1 j ∈ V2
i∈V1

y binary.
Location-covering problems

Greedy algorithm (Chvatal 1979)

1 Set N = V2 , I = ∅
2 While N ̸= ∅ do
P
1 For every facility i ∈ V1 \ I compute βiN = j∈N aij
∗
2 Select the facility i ∈ V1 \ I that maximizes βiN /fi
3 Update I ← I ∪ {i∗ }, N ← N \ {j ∈ V2 : ai∗ j = 1}
p-Centre problems (pCP)

There are situations where it may impossible to satisfy the covering

condition for every customer node, for instance due to budget
considerations
In this case, we may try to answer an alternative question: If I have a
budget to locate up to p facilities, where should I do so to minimize
the largest distance from any customer to its closest facility?
In the discrete variant, we usually assume that the set of potential
facilities V1 coincides with the set of customer locations V2 ,
therefore we let V1 = V2 = V
The distances are usually denoted dij
Binary search for the pCP

Main idea
Let r > 0 be a target radius. Construct a set-covering matrix A such
that for every i, j ∈ V , aij = 1 if dij < r. Let fi = 1 and solve the
associated set-covering problem. An optimal value ≤ p indicates the
possibility of finding a solution to the pCP of value r
Binary search for the pCP

The following method solves the pCP for integer distances dij , i, j ∈ V

Binary search
1 Let rL = 0, rU = max{dij , i, j ∈ V }
2 While rL < rU do
1 Let r = ⌊ rL +r2
U
⌋
2 Solve the associated set-covering problem
3 If feasible, let rU ← r
4 Else, let rL ← r + 1
3 Return the solution associated to the last value of rU
Two-stage stochastic location models

To fix ideas, let us consider a single-commodity capacitated location

problem. Decisions typically occur at two levels
Strategic: Where to locate the facilities
Tactical: How to allocate demand points to facilities

Two-stage stochastic location

It is natural to pose the problem as a two-stage stochastic optimization
problem. At the first level, location decisions must be taken while taking
into consideration that demands and transportation costs may be
random. Then, as demand and costs are revealed, allocation takes place
as a second level decision (recourse)
Two-stage stochastic location models

Uncertainty is modeled via a set S of scenarios. Each scenario s is

defined by:
A demand vector djs , j ∈ V2 for every customer j ∈ V2
A unit cost matrix cijs , i ∈ V1 , j ∈ V2
A probability ps of the scenario to be observed

First-level decision variables

yi ∈ {0, 1}, i ∈ V1 indicating whether facility i is placed

Second-level decision variables

0 ≤ xijs ≤ 1, i ∈ V1 , j ∈ V2 , s ∈ S indicating the fraction of the demand
of customer j that is serviced by facility i, at scenario s
Multi-commodity two-echelon location

We can formulate this problem as follows:

X X
Minimize Z = fi yi + ps cijs xijs
x,y
i∈V1 i∈V1 ,j∈V2 ,s∈S

subject to
X
xijs = 1 j ∈ V2 , s ∈ S
i∈V1
X
djs xijs ≤ qi yi i ∈ V1 , s ∈ S
j∈V2

x≥0
y binary.