Optimización
O ti i ió Avanzada
A d
Principio de Descomposición de
Dantzig-Wolfe
Oct. 12, 2006
Andrés Medaglia, Ph.D.
Profesor Asociado
Departamento de Ingeniería Industrial
Centro de Optimización y Probabilidad Aplicada
(http://copa.uniandes.edu.co)
Fuente: S. C. Fang – Lecture Notes on Large Scale Optimization, NC State U.
Decomposition principle
m constraints
Andrés L. Medaglia, Ph.D.
Decomposition principle
bottleneck
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
How many variables and
contraints?
Andrés L. Medaglia, Ph.D.
Decomposition principle
(LP) ((MP))
LP MP
variables ? ?
constraints ? ?
Andrés L. Medaglia, Ph.D.
Decomposition principle
(LP) ((MP))
LP MP How many
H
variables n too many is too
many?
constraints m0+m1=m m0+1
Andrés L. Medaglia, Ph.D.
Decomposition principle
How many is too many?
As many
y as vertices in X
Andrés L. Medaglia, Ph.D.
Decomposition principle
How to solve MP?
(MP)
Note: MP has a huge
g number of variables.
Andrés L. Medaglia, Ph.D.
Column generation
(MP)
Suppose we have a bfs for MP:
The dual variables are:
Andrés L. Medaglia, Ph.D.
Column generation
t i l column
typical l off MP
( )
(MP)
Is the current solution optimal?
p
Andrés L. Medaglia, Ph.D.
Column generation
Lets check the reduced costs for the
nonbasic variables
∀q ∈ N% ( = I N )
Andrés L. Medaglia, Ph.D.
Column generation
Optimality condition:
∀q ∈ N% ( = I N )
⇒ ∀q ∈ N% ( = I N )
Andrés L. Medaglia, Ph.D.
Column generation
How to reconstruct the solution in the
original variables?
Andrés L. Medaglia, Ph.D.
Column generation
How to reconstruct the solution in the
original variables?
Andrés L. Medaglia, Ph.D.
Column generation
What if the optimality conditions are not
met?
if ∃q ∈ N% ( = I N ) such that f q < 0
the q-th variable enters the basis of MP.
generated (entering) column
Andrés L. Medaglia, Ph.D.
Column generation
optimality conditions:
⇒ ∀q ∈ N% ( = I N )
Andrés L. Medaglia, Ph.D.
Column generation problem
Andrés L. Medaglia, Ph.D.
Column generation problem
Andrés L. Medaglia, Ph.D.
Column generation problem
optimality conditions:
⇒ ∀q ∈ N% ( = I N )
Andrés L. Medaglia, Ph.D.
Column generation problem
optimality conditions:
⇒ ∀q ∈ N% ( = I N )
optimality conditions revisited:
Andrés L. Medaglia, Ph.D.
Column generation problem
What if the optimality conditions are not
met?
generated (entering) column
Andrés L. Medaglia, Ph.D.
Decomposition principle: summary
x*
Andrés L. Medaglia, Ph.D.
Ejemplo: múltiples plantas
Una compañía con 2 plantas (A y B) y 2
productos (estándar y premium).
C d producto
Cada d t requierei unas horas
h de
d torno
t y
otras horas de fresa.
Hay límites de horas por proceso y por planta.
La planta A tiene mejor tecnología que la planta
B, es decir requiere de menos horas de torno y
f
fresa para producir
d i una unidad.
id d
La compañía tiene un total de 120 kg de materia
prima (acero especial) y cada unidad requiere
de 4kg de acero.
Andrés L. Medaglia, Ph.D.
Ejemplo: múltiples plantas
Planta
Pl t A Pl t B
Planta
Est. Prem. Est. Prem.
Torno 4 2 5 3
Fresa 2 5 5 6
Planta
A B Est. Prem.
TTorno 80 60 B
Beneficio
fi i 10 15
Fresa 60 75
Andrés L. Medaglia, Ph.D.
Ejemplo: múltiples plantas
xij : cantidad de producto j a producir en planta i
i = {1,2}, j = {1,2}
planta A planta B estándar premium
max 10 x11 + 15 x12 + 10 x21 + 15 x22
s.a.,
4 x11 + 4 x12 + 4 x21 + 4 x22 ≤ 120
4 x11 + 2 x12 ≤ 80
2 x11 + 5 x12 ≤ 60
5 x21 + 3 x22 ≤ 60
5 x21 + 6 x22 ≤ 75
xij ≥ 0
Andrés L. Medaglia, Ph.D.
Ejemplo: múltiples plantas
max 10 x11 + 15 x12 + 10 x21 + 15 x22
sa
s.a.,
4 x11 + 4 x12 + 4 x21 + 4 x22 ≤ 120
4 x11 + 2 x12 ≤ 80
2 x11 + 5 x12 ≤ 60
5 x21 + 3 x22 ≤ 60
5 x21 + 6 x22 ≤ 75
xij ≥ 0
Andrés L. Medaglia, Ph.D.
Ejemplo: múltiples plantas
max 10 x11 + 15 x12 + 10 x21 + 15 x22
sa
s.a.,
4 x11 + 4 x12 + 4 x21 + 4 x22 ≤ 120
4 x11 + 2 x12 ≤ 80
2 x11 + 5 x12 ≤ 60
5 x21 + 3 x22 ≤ 60
5 x21 + 6 x22 ≤ 75
xij ≥ 0
Andrés L. Medaglia, Ph.D.
Diagonal block structure
¿Cómo
Có resolvemos
l un programa lineal
li l de
d gran escala
l
con estructura diagonal en bloques?
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 -1
1
-1 1 -1
-1 1
-1 1 1
1 -1
-1 1 -1
-1 1
-1
11 1
1 -1
-1 1 -1
-1 1
-1 1 1
Andrés L. Medaglia, Ph.D.
Special diagonal block structure in A1
Andrés L. Medaglia, Ph.D.
Special diagonal block structure of
Multicommodity network flows
Andrés L. Medaglia, Ph.D.
Diagonal block structure
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
Andrés L. Medaglia, Ph.D.
Decomposition principle
optimality conditions revisited for block structure:
Andrés L. Medaglia, Ph.D.
Column generation
What if the optimality conditions are not
met?
generated (entering) column
Andrés L. Medaglia, Ph.D.
Other types of structures
IIs it possible
ibl to
t apply
l the
th decomposition
d iti approach
h to
t the
th
following problem?
A 0 B1 b1
A1 B 2 b2
A 2 B3 b3
O
A k −1 B k bk
Andrés L. Medaglia, Ph.D.
Other types of structures
S l ti
Solution approach...
h
A 0 B1 b1
A1 B 2 b2
A 2 B3 b3
O M
A k − 2 B k −1 b k −1
A k −1 B k bk
Andrés L. Medaglia, Ph.D.
Other types of structures
S l ti
Solution approach...
h
A1 B 2 b2
O M
A k − 2 B k −1 b k −1
A 0 B1 b1
A 2 B3 b3
O M
A k −1 B k bk
Andrés L. Medaglia, Ph.D.
Other types of structures
IIs it possible
ibl to
t apply
l the
th decomposition
d iti approach
h to
t the
th
following problem?
A1
mon columnns
b1
A2 b2
A1 b3
Comm
O
Ak bk
Andrés L. Medaglia, Ph.D.
Other types of structures
IIs it possible
ibl to
t apply
l the
th decomposition
d iti approach
h to
t the
th
following problem?
A0 b0
A11 b1
A12 b2
A10
A13 b3
O M
A1k bk
Andrés L. Medaglia, Ph.D.
