Multiobjectivo

Introduction
Solution Methods
Applications
Conclusion
Multiobjective Combinatorial Optimization

New Trends in Methods and Aplications
Matthias Ehrgott
Department of Engineering Science

The University of Auckland, New Zealand
Operations Research 2005, Bremen, 7. – 9. September 2005
Matthias Ehrgott Multiobjective Combinatorial Optimization

Introduction
Solution Methods
Applications
Conclusion
with Xavier Gandibleux, Laboratoire d’Informatique de Nantes

Atlantique, Université de Nantes

Introduction
Solution Methods
Applications
Conclusion
Overview
1 Introduction
Formulation
Definitions of Optimality
Characteristics
2 Solution Methods
Scalarization
The Two Phase Method
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Solution Methods
Applications
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Solution Methods
Applications
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Solution Methods
Applications
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Mathematical Formulation
min z(x) = Cx
subject to Ax = b
x ∈ {0, 1}n
x ∈ {0, 1}n −→ n variables, i = 1, . . . , n

C ∈ Zp×n −→ p objective functions, k = 1, . . . , p
A ∈ Zm×n −→ m constraints, j = 1, . . . , m

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
min z(x) = Cx
subject to Ax = b
x ∈ {0, 1}n
x ∈ {0, 1}n −→ n variables, i = 1, . . . , n


Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
min z(x) = Cx
subject to Ax = b
x ∈ {0, 1}n
x ∈ {0, 1}n −→ n variables, i = 1, . . . , n


Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Feasible Sets
10
X = {x ∈ {0, 1}n : Ax = b} 7
feasible set in decision space 6
Y = z(X ) = {Cx : x ∈ X } 5
feasible set in objective space 4
conv(Y ) + Rp= 3
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Feasible Sets
10
X = {x ∈ {0, 1}n : Ax = b} 7
Y = C(X)
Y = z(X ) = {Cx : x ∈ X } 5
conv(Y ) + Rp= 3
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Feasible Sets
10
8
conv(C(X)) + Rp=
X = {x ∈ {0, 1}n : Ax = b} 7
Y = z(X ) = {Cx : x ∈ X } 5
conv(Y ) + Rp= 3
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Lexicographic Optimality
10
Individual minima 9
zk (x̂) ≤ zk (x) for all x ∈ X 8
Lexicographic optimality (1) 6
z(x̂) ≤lex z(x) for all x ∈ X 5
z π (x̂) ≤lex z π (x) for all x ∈ X 3
and some permutation z π of 2
(z1 , . . . , zp ) 1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
10
Individual minima 9
(z1 , . . . , zp ) 1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
10
Individual minima 9
(z1 , . . . , zp ) 1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10
Weakly efficient solutions XwE

9
There is no x with z(x) < z(x̂)

8
componentwise
7
z(x̂) is weakly nondominated

6
YwN := z(XwN )
5
Efficient solutions XE 4
There is no x with z(x) ≤ z(x̂) 3
and z(x) 6= z(x̂) 2
z(x̂) is nondominated 1
YN := z(XE ) 0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10
Weakly efficient solutions XwE

9
There is no x with z(x) < z(x̂)

8
componentwise
7
z(x̂) is weakly nondominated

6
YwN := z(XwN )
5
Efficient solutions XE 4
There is no x with z(x) ≤ z(x̂) 3
and z(x) 6= z(x̂) 2
z(x̂) is nondominated 1
YN := z(XE ) 0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10
Supported efficient solutions

9
XSE : There is λ > 0 with

λT C x̂ ≤ λT Cx for all x ∈ X 8
7
C x̂ is extreme point of
conv(Y ) + Rp= → XSE 1 6
C x̂ is in relative interior of 5
face of conv(Y ) + Rp= → XSE 2 4
Nonsupported efficient solutions 2
XNE : C x̂ is in interior of
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10

9

conv(C(X)) + Rp=
7
5
C x̂ is in relative interior of
5 1

3
λ= ,
6 6
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10

9

conv(C(X)) + Rp=
7
5
XNE : C x̂ is in interior of 1 1

λ= 2, 2
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10

9

conv(C(X)) + Rp=
7
5

λ= 3, 3
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10

9

conv(C(X)) + Rp=
7
5

λ= 3, 3
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Efficient Solutions
10

9

conv(C(X)) + Rn=
7
5
conv(Y ) + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Classification of Efficient Sets
Hansen 1979:
x 1 , x 2 ∈ XE are equivalent if
Cx 1 = Cx 2 XE
Complete set: X̂ ⊂ XE such
that for all y ∈ YN there is XSE XN E
x ∈ X̂ with z(x) = y
Minimal complete set contains XSE1 XEm XSE2 XN E m
no equivalent solutions
XSE1m XSE2m
Maximal complete set contains
all equivalent solutions

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Hansen 1979:
Cx 1 = Cx 2 XE
XSE1m XSE2m

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Hansen 1979:
Cx 1 = Cx 2 XE
XSE1m XSE2m

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Hansen 1979:
Cx 1 = Cx 2 XE
XSE1m XSE2m

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
MOCO Problems Are Hard
Theorem
Multiobjective combinatorial optimization problems are NP-hard,
#P-complete, and intractable.
Examples:
Shortest path (Hansen 1979, Serafini 1986)
Assignment (Serafini 1986, Neumayer 1994)
Spanning tree (Hamacher and Ruhe 1994)
Network flow (Ruhe 1988)

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Theorem
Examples:

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Theorem
Examples:

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Theorem
Examples:

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Theorem
Examples:

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Number of Efficient Solutions
Intractable: XE , even YSN , can be exponential in the size of the

instance
Empirically often
|XNE | grows exponentially with instance size
|XSE | grows polynomially with instance size
But this depends on numerical values of C

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion

instance
Empirically often

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion

instance
Empirically often

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion

instance
Empirically often

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion

instance
Empirically often

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Bi-KP Problems : proportion de supportees

0.3
A
B
C
D
0.25
0.2
nb SE par rap. NE
0.15
0.1
0.05
0
50 100 150 200 250 300 350 400 450 500
instance

Introduction
Formulation
Solution Methods
Optimality
Applications
Characteristics
Conclusion
Bi-AP Problems : prop SE par rap. NE

1
A
B
C
0.9 D
0.8
prop SE par rap. NE
0.7
0.6
0.5
0.4
0.3
0.2
05 10 15 20 25 30 35 40 45 50
instance

Introduction Scalarization
Solution Methods The Two Phase Method
Applications Branch and Bound
Conclusion Metaheuristics
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Principle and Properties of Scalarization
Convert multiobjective problem to (parameterized) single objective

problem and solve repeatedly with different parameter values
Desirable properties of scalarizations: (Wierzbicki 1984)

Correctness: Optimal solutions are (weakly) efficient
Completeness: All efficient solutions can be found
Computability: Scalarization is not harder than single
objective version of problem (theory and practice)
Linearity: Scalarization has linear formulation
















Scalarization Methods
10
Weighted
n sum:
o 8
T
min λ z(x) 7
x∈X
ε-constraint: 6
min {zl (x) : zk (x) ≤ εk , k 6= l} 5
x∈X 4
Weighted
Chebychev: 3
I
min max νk (zk (x) − yk ) 2
x∈X k=1,...,p 3 4

1
λT = 7, 7
0
0 1 2 3 4 5 6 7 8 9 10

10
Weighted
n sum:
o z1(x) ≤ 5.5
8
T
min λ z(x) 7
x∈X
ε-constraint: 6
min {zl (x) : zk (x) ≤ εk , k 6= l} 5
x∈X 4
Weighted
Chebychev: 3
I
x∈X k=1,...,p 1
0
0 1 2 3 4 5 6 7 8 9 10

10
Weighted
n sum:
o 8
T
min λ z(x) 7
x∈X
ε-constraint: 6
min {zl (x) : zk (x) ≤ εk , k 6= l} 5
x∈X 4
Weighted Chebychev: 3
ν2 = 1

I
x∈X k=1,...,p 1
0
yI ν1 = 0.5
0 1 2 3 4 5 6 7 8 9 10

General Formulation
p
( )
p X
min max [νk (ck x − ρk )] + [λk (ck x − ρk )]
x∈X k=1
k=1
subject to c k x ≤ εk k = 1, . . . , p
Includes Correct Complete Computable Linear

Weighted sum + - + +
ε-constraint + + - +
Benson + + - +
Chebychev + (+) (-) +
Max-ordering + + - +
Reference point + (+) (-) +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
p
( )
p X
x∈X k=1
k=1

Benson + + - +

General Formulation
Theorem (Ehrgott 2005)

1 The general scalarization is NP-hard.
2 An optimal solution of the Lagrangian dual of the linearized
general scalarization is a supported efficient solution.

Method of Elastic Constraints
10
8
z1(x) ≤ 5.5
X 7
min cl x + µk wk 6
x∈X
k6=l 5
s.t. ck x + vk − wk ≤ εk k 6= l 4
vk , w k ≥ 0 k 6= l 3
0
0 1 2 3 4 5 6 7 8 9 10

10
8
z1(x) ≤ 5.5
X 7
min cl x + µk wk 6
x∈X
k6=l 5
µ1 = 1.5
s.t. ck x + vk − wk = εk k 6= l 4
vk , w k ≥ 0 k 6= l 3
0
0 1 2 3 4 5 6 7 8 9 10

Theorem (Ehrgott and Ryan 2002)

The method of elastic constraints
is correct and complete,
contains the weighted sum and ε-constraint method as special
cases,
is NP-hard.
... but (often) solvable in practice because

it “respects” problem structure
it “limits damage” of ε-constraints


cases,
is NP-hard.



cases,
is NP-hard.



cases,
is NP-hard.


Integer Programming Duality
Theorem (Klamroth et al. 2004)

x̂ ∈ XE if and only if there is
F̂ ∈ F := {F : Rm+p−1 → R nondecreasing}
such that x̂ is an optimal solution to
n o
max cj x − F̂ ((ck x)k6=j , b) : Ax 5 b, x = 0, x integer .
F̂ can
n be chosen as an optimal solution of the IP dual o
min F (−e, b) : F ((−ck x)k6=j , Ax) ≥ cj x ∀x ∈ Zn= , F ∈ F
of max{cl x : ck x ≥ εk , k 6= l, Ax = b, x ∈ Zn= }
The level curve of the objective function of the composite IP
at level 0 defines an upper bound on YN .

Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Phase 1: Compute XSE

1 Find lexicographic solutions
2 Recursively:
Calculate λ
Solve min λT Cx
x∈X
Phase 2: Compute XNE
1 Solve by triangle
2 Use neighborhood (wrong)
3 Use constraints (bad)
4 Use variable fixing (possible)
5 Use ranking (good)

Phase 1: Compute XSE 10
1 Find lexicographic solutions 9
2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3
3 Use constraints (bad) 2
4 Use variable fixing (possible) 1
5 Use ranking (good) 0

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

2 Recursively: 8
Calculate λ 7
Solve min λT Cx 6
x∈X
5

4
1 Solve by triangle
3

0 1 2 3 4 5 6 7 8 9 10

Enumeration Problems
Finding maximal complete set:

Enumeration to find all optimal solutions of minx∈X λT Cx
Enumeration to find all x ∈ XNE with Cx = y ∈ YND
Finding minimal complete set:
Enumeration to find XSE 2









235
z(x) ∈ Z SN1
( 199,230 )
230
z(x) ∈ Z SN2
( 202,227 )
( 203,226 )
225 ( 204,225 )
( 206,223 )
λ = 0.5666 ( 207,222 )
( 208,221 )
( 209,220 )
z2
220 ( 210,219 )
( 211,218 )
( 212,217 )
λ =0.5 ( 213,216 )
( 214,215 )
215
( 215,214 )
( 216,213 )
( 218,211 )
210 ( 220,209 )
λ =0.4918
205
195 200 205 210 215 220 225
z1

2 Phase Algorithm for Biobjective Assignment

Przybylski et al. 2004:
Hungarian Method for minx∈X λT Cx
Enumeration of all optimal solutions of minx∈X λT Cx
(Fukuda and Matsui 1992)
Ranking of (non-optimal) solutions of minx∈X λT Cx
(Chegireddy and Hamacher 1987)
Results for 100 × 100:
Range Variable Fixing Seek & Cut Ranking
20 14049.17 2755.75 220.07
40 × 17441.35 225.06
60 × 38553.18 399.65
80 × 53747.45 721.08
100 × 60227.31 711.97


20 14049.17 2755.75 220.07
40 × 17441.35 225.06
60 × 38553.18 399.65
80 × 53747.45 721.08
100 × 60227.31 711.97


20 14049.17 2755.75 220.07
40 × 17441.35 225.06
60 × 38553.18 399.65
80 × 53747.45 721.08
100 × 60227.31 711.97


20 14049.17 2755.75 220.07
40 × 17441.35 225.06
60 × 38553.18 399.65
80 × 53747.45 721.08
100 × 60227.31 711.97


20 14049.17 2755.75 220.07
40 × 17441.35 225.06
60 × 38553.18 399.65
80 × 53747.45 721.08
100 × 60227.31 711.97

Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Multicriteria Branch and Bound
Ulungu and Teghem 1997, N0
Mavrotas and Diakoulaki 2002 ∅

∅
Branching: As in single
objective case N1
1, 2, 3
N2
1, 2
N5
1
N9
∅
∅ 3 2 1
Bounding: Ideal point of L=

23
27

problem at node is dominated N3 N4 N6 N7

1, 2, 4 1, 2 1, 3 1
by efficient solution 3 3, 4 2 2, 3
23 29

Branching may be very L= 27 , 16
N8
ineffective 1, 4, 5
2, 3
Use lower and upper bound sets L=

23

27, 29
16

, 27

18


∅
objective case N1
1, 2, 3
N2
1, 2
N5
1
N9
∅
∅ 3 2 1

23
27


1, 2, 4 1, 2 1, 3 1
23 29

N8
ineffective 1, 4, 5
2, 3

23

27, 29
16

, 27

18


∅
objective case N1
1, 2, 3
N2
1, 2
N5
1
N9
∅
∅ 3 2 1

23
27


1, 2, 4 1, 2 1, 3 1
23 29

N8
ineffective 1, 4, 5
2, 3

23

27, 29
16

, 27

18


∅
objective case N1
1, 2, 3
N2
1, 2
N5
1
N9
∅
∅ 3 2 1

23
27


1, 2, 4 1, 2 1, 3 1
23 29

N8
ineffective 1, 4, 5
2, 3

23

27, 29
16

, 27

18


∅
objective case N1
1, 2, 3
N2
1, 2
N5
1
N9
∅
∅ 3 2 1

23
27


1, 2, 4 1, 2 1, 3 1
23 29

N8
ineffective 1, 4, 5
2, 3

23

27, 29
16

, 27

18

Bound Sets
Ehrgott and Gandibleux 2005:

1 Lower bound set L
10
is Rp= -closed 9
is Rp= -bounded 8
YN ⊂ L + Rp= 7

L ⊂ L + Rp=
6
N 5
2 Upper bound set U 4
is Rp= -closed 3
is Rp= -bounded 2
h c i
YN ∈ cl U + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10
U ⊂ U + Rp=
N

Bound Sets
Ehrgott and Gandibleux 2005:

1 Lower bound set L
10
is Rp= -closed 9
is Rp= -bounded 8
YN ⊂ L + Rp= 7

L ⊂ L + Rp=
6
N 5
2 Upper bound set U 4
is Rp= -closed 3
is Rp= -bounded 2
h c i
YN ∈ cl U + Rp=
1
0
0 1 2 3 4 5 6 7 8 9 10
U ⊂ U + Rp=
N

90000
ub ?
lb r
80000 dMax e
r?
70000 ?r
?
60000 ?
?
?r
y2 50000 ?
?
?
40000
rr
rr ?
rrr ???e
30000 r e ??????r
rrr
????
?? r? ?
20000 rrrrrrr ?rrrr??r?? ? ? ?? ?
r ? ?r
10000
10000 20000 30000 40000 50000 60000 70000 80000 90000
y1

Approximation Algorithms
Approximation algorithms with approximation ratio 1 + ε find

ε-efficient solutions
FPTAS for shortest path (Warburton 1987, Hansen)
PTAS for XSE of network flow (Ruhe and Fruhwirth 1990)
FPTAS for knapsack (Erlebach et al. 2002)
Travelling salesman (Ehrgott 2000, Angel et al. 2004)
Scheduling (Angel et al. 2005 )
Existence of approximation algorithms (Safer 1992,
Papadimitriou and Yannakakis 2000)













Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Heuristics versus Metaheuristics
Heuristic: Technique which seeks near-optimal solutions at a

reasonable computational cost without being able to
guarantee optimality ... often problem specific
(Reeeves 1995)
Examples: Multiobjective greedy and local search (Ehrgott and
Gandibleux 2004, Paquete et al. 2005)
Metaheuristic: Iterative master strategy that guides and modifies
the operations of subordinate heuristics by combining
intelligently different concepts for exploring and
exploiting the search space ... applicable to a large
number of problems. (Glover and Laguna 1997,
Osman and Laporte 1996)
Examples: MOEA, MOTS, MOSA etc.


(Reeeves 1995)


(Reeeves 1995)


(Reeeves 1995)

The Quality of Heuristics
z2
z1
Which approximation is best?

Cardinal and geometric measures

(Hansen and Jaszkiewicz 1998)
Hypervolumes (Zitzler and Thiele 1999)
Coverage, uniformity, cardinality (Sayin 2000)
Distance based measures (Viana and de Sousa 2000)
Integrated convex preference (Kim et al. 2001)
Volume based measures (Tenfelde-Podehl 2002)
Analysis and Review (Ziztler et al. 2003)













Timeline for Multiobjective Metaheuristics
Evolutionary Algorithms
1984: VEGA by Schaffer
Neural Networks
1990: Malakooti
Neighbourhood Search Algorithms
1992: Simulated Annealing by Serafini
1992/93: MOSA by Ulungu and Teghem
1996/97: MOTS by Gandibleux et al.
1997: TS by Sun
1998: GRASP by Gandibleux et al.
Hybrid and Problem Dependent Algorithms
1996: Pareto Simulated Annealing by Czyzak and Jaszkiewicz
1998: MGK algorithm by Morita et al.
Many more since 2000

Neural Networks
1990: Malakooti
1997: TS by Sun

Neural Networks
1990: Malakooti
1997: TS by Sun

Neural Networks
1990: Malakooti
1997: TS by Sun

Multiobjective Evolutionary Algorithms
Main principles
1 Population of solutions
2 Self adaptation (independent evolution)
3 Cooperation (exchange of information)
Main problems
1 Uniform convergence (fitness assignment by ranking and
selection with elitism)
2 Uniform distribution (niching, sharing)
Huge number of publications, including surveys, books, EMO
conference series
Few applications to MOCO problems

Main principles
Main problems
conference series

Main principles
Main problems
conference series

Main principles
Main problems
conference series

Multiobjective Neighbourhood Search Algorithms
Multiobjective Simulated Annealing Multiobjective Tabu Search

Initial solution Initial solution
Neighbourhood structure Neighbourhood structure
Scalarizing function s(z(x), λ) Scalarizing function s(z(x), λ)
Set of weights (directions) λ Set of weights (directions) λ
Simulated annealing based on s Tabu search based on s
Merge sets of solutions Merge sets of solutions











Hybrid Algorithms
1 EA components in NSA (to improve coverage)

Pareto Simulated Annealing (Czyzak and Jaszkiewicz 1996)
Use set of starting solutions
SA uses information from set of solutions
TAMOCO (Hansen 1997, 2000)
TS uses information from set of solutions
2 NSA strategies inside EAs (to improve convergence)
MGK (Morita et al. 1998, 2001)
Use local search to improve solutions
MOGLS (Jaszkiewicz 2001)

Hybrid Algorithms


Hybrid Algorithms


Hybrid Algorithms


Hybrid Algorithms


Hybrid Algorithms


MOMKP n=250, p=2, k=2
10500
E
VEGA
10000
9500
9000
8500
8000
7500
7000 7500 8000 8500 9000 9500 10000
z1

MOMKP n=250, p=2, k=2
10500
E
VEGA
SPEA
10000 NSGA
9500
9000
8500
8000
7500
7000 7500 8000 8500 9000 9500 10000
z1

MOMKP n=250, p=2, k=2
10500
E
VEGA
SPEA
NSGA
10000 MOGLS
MOGTS
9500
9000
8500
8000
7500
7000 7500 8000 8500 9000 9500 10000
z1

MOMKP n=250, p=2, k=2
10500
E
VEGA
SPEA
NSGA
MOGLS
10000
MOGTS
MGK
9500
9000
8500
8000
7500
7000 7500 8000 8500 9000 9500 10000
z1

Hybrid Algorithms
3 Alternate between EA and NSA

Ben Abdelaziz et al. 1997, 1999
Use GA to find first approximation
Apply TS to improve reults of GA
Delorme et al. 2003, 2005
Use Use GRASP to compute first approximation
Use SPEA to improve results of GRASP

Hybrid Algorithms


Hybrid Algorithms



Hybrid Algorithms
4 Combination of EA, NSA and problem dependent components
Gandibleux et al. 2003, 2004

Use crossover and population as EA component
Use path-relinking as NSA component
Use XSEm and bound set as problem dependent component

Hybrid Algorithms


Hybrid Algorithms

x1 x2
swap
initiating move move guiding

solution neighbor neighbor solution

Hybrid Algorithms
5 Combination of heuristics and exact algorithms

Gandibleux and Fréville 2000
Use tabu search as heuristic
Use cuts to eliminate search areas where (provably) no efficient
solutions exist
Przybylski et al. 2004
Use two phase method as exact algorithm
Use heuristic to reduce search space

Hybrid Algorithms

solutions exist

Hybrid Algorithms

solutions exist

Results with the bound 2

16000
2ph
(PR1)(N1)
(PR1)(N2)
14000 (PR2)(N1)
(PR2)(N2)
(PR3)(N1)
12000
10000
CPU_t (s)
8000
6000
4000
2000
0
50 55 60 65 70 75 80 85 90 95 100
instance size

New Metaheuristic Schemes for MOCO
1 Ant colony optimization (Iredi et al. 2001, Gravel et al. 2002,

Doerner et al. since 2001)
2 Scatter search (Beausoleil 2001, Molina 2004)
3 Particle swarm (Coello 2002, Mostaghim 2003)
4 Constraint programming (Barichard 2003)







Introduction Finance
Solution Methods Transportation
Applications Medicine
Conclusion Telecommunication
Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Portfolio Selection
Markowitz 1952 with cardinality constraint, e.g. Chang et al. 2000
max z1 (x) = µT x 350
T
min z2 (x) = x σx 300
250
subject to e T x = 1 200
f1
xi ≤ ui yi 150
100
xi ≥ li yi 50
eT y = k 0
0 200 400 600 800 1000 1200 1400 1600
n f2
y ∈ {0, 1}

Portfolio Selection
Markowitz 1952 with cardinality constraint, e.g. Chang et al. 2000
max z1 (x) = µT x 350
T
min z2 (x) = x σx 300
250
subject to e T x = 1 200
f1
xi ≤ ui yi 150
100
xi ≥ li yi 50
eT y = k 0
0 200 400 600 800 1000 1200 1400 1600
n f2
y ∈ {0, 1}

Other Problems in Portfolio Selection
Additional objectives e.g. Steuer and Na 2003

First International Workshop on Multiattribute Portfolio
Selection Helsinki 2005
“full potential of multiple criteria decision making (MCDM)
applications in portfolio management has not been unleashed”
Chapter 20 in Figueira, Greco, Ehrgott “Multicriteria Decision
Analysis” Springer 2005
Project Portfolio Selection (Doerner et al. 2004)







Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Railway Capacity Optimization
Assign trains to tracks in railway

node to
maximize number of trains that
can pass node
minimize disruption in case of
delays
Bicriteria set packing problem
GRASP (Delorme et al. 2002)
Branch and Bound (Ryan and
Ehrgott 2005)


node to
can pass node
delays
Ehrgott 2005)


node to
can pass node
delays
Ehrgott 2005)


node to
can pass node
delays
Ehrgott 2005)


node to
can pass node
delays
Ehrgott 2005)


node to
can pass node
delays
Ehrgott 2005)

Airline Crew Scheduling

Partition flights into set of pairings, but minimizing cost can cause
delays ...
and be very expensive


Partition flights into set of pairings, but minimizing cost can cause
delays ...
and be very expensive
Sunday, 4 August, 2002, 20:29 GMT 21:29 UK
Delays as Easyjet cancels 19 flights
Passengers with low-cost airline Easyjet are suffering delays after 19 flights in and out
of Britain were cancelled.
The company blamed the move - which comes a week after passengers staged a protest sit-in
at Nice airport - on crewing problems stemming from technical hitches with aircraft.
Crews caught up in the delays worked up to their maximum hours and then had to be allowed
home to rest.
Mobilising replacement crews has been a problem as it takes time to bring people to
airports from home. Standby crews were already being used and other staff are on holiday.


Model 1: Minimize cost and minimize non-robustness (Ehrgott and
Ryan 2002)

1 pairing j includes flight i
aij =
0 otherwise
min z1 (x) = c T x
min z2 (x) = r T x
subject to Ax = e
Mx = b
x ∈ {0, 1}n


Ryan 2002)

aij =
0 otherwise
min z1 (x) = c T x
min z2 (x) = r T x
subject to Ax = e
Mx = b
x ∈ {0, 1}n


Ryan 2002)

aij =
0 otherwise
min z1 (x) = c T x
min z2 (x) = r T x
subject to Ax = e
Mx = b
x ∈ {0, 1}n

6000
LP
IP 2% bound gap
Cost IP solution 2% bound gap
5000
4000
Non-robustness
3000
2000
1000
0
880 900 920 940 960 980 1000 1020 1040
Cost

Model 2: Minimize cost, maximize unit crewing (Tam et al. 2004)
min c1T x1 + c2T x2

min e T s1 + e T s2
subject to A1 x1 = e
M1 x 1 = b1
A2 x2 = e
M2 x2 = b2
U1 x1 − U2 x2 − s1 + s2 = 0
x1 x2 s1 s2 ∈ {0, 1}n

Model 2: Minimize cost, maximize unit crewing (Tam et al. 2004)
min c1T x1 + c2T x2

min e T s1 + e T s2
subject to A1 x1 = e
M1 x 1 = b1
A2 x2 = e
M2 x2 = b2
U1 x1 − U2 x2 − s1 + s2 = 0
x1 x2 s1 s2 ∈ {0, 1}n


Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Radiotherapy Treatment Design
Choose beam directions and intensities to destroy tumour and

spare healthy organs (e.g. Holder 2004)
min(zT , zS , zN )
subject to AT x + zT e ≥ lT
AT x ≤ uT
AS x − zS e ≤ uS
AN x − zN e ≤ uN
zS ≥ −uS
zN ≥ 0
x ≥ 0
x ≤ Mye
y ∈ {0, 1}n


min(zT , zS , zN )
subject to AT x + zT e ≥ lT
AT x ≤ uT
AS x − zS e ≤ uS
AN x − zN e ≤ uN
zS ≥ −uS
zN ≥ 0
x ≥ 0
x ≤ Mye
y ∈ {0, 1}n


min(zT , zS , zN )
subject to AT x + zT e ≥ lT 50
≤
45
AT x uT 40
AS x − zS e ≤ uS 35
AN x − zN e ≤ uN 30
zS ≥ −uS 25
≥
20
zN 0 15
x ≥ 0 10
x ≤ Mye 5
y ∈ {0, 1}n 5 10 15 20 25 30 35 40 45 50

Overview
1 Introduction
Formulation
Characteristics
2 Solution Methods
Scalarization
Branch and Bound
Metaheuristics
3 Applications
Finance
Transportation
Medicine
Telecommunication
4 Conclusion
Routing in IP Networks
EUROPE2
2500
2 168 901
2000
3 4 568 033
Beugnies et al. 2003 2250

6 977 311 500
4 641 359
750
5 13
Find routes r with objectives 2 250
8 453 025
1 099 044
250 1500
6 16 2 248 390
min z1 (r ) = (i,j)∈r c 1 (i, j)

P 3 579 607
500
1
420 644
delay 2250
1750
4 024 195
500
1 695 546
1 915 678 8
max z2 (r ) = min(i,j)∈r c 2 (i, j) 9

250
226 621 2500
456 491
15
bandwidth 2250
1000
372 461
443 005
min z3 (r ) = |{(i, j) ∈ r }|
2000
429 359
14 10
250
number of hops 11
819 462
4750
1500
478 379
12 780 006
Euronet PoP durée (µs)

BPR (Kb/s)
Non PoP node

Other Problems in Telecommunication
Bandwidth allocation (Ogryczak et al. 2005)

Placement of internet caches (Wierzbicki 2002)
Network design


Network design


Network design


Network design

Introduction
Solution Methods
Applications
Conclusion
Topics for Research
Representative set of efficient solutions

Consideration of DM preferences
Quality of approximation with heuristics
Better use of knowledge in combinatorial optimization
Algorithms that combine heuristics, exact methods, problem
specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Topics for Research

specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Topics for Research

specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Topics for Research

specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Topics for Research

specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Topics for Research

specific methods
New applications

Introduction
Solution Methods
Applications
Conclusion
Some References
M. Ehrgott, X. Gandibleux. Multiobjective Combinatorial
Optimization. In Multiple Criteria Optimization. Kluwer 2002.
M. Ehrgott, X. Gandibleux, J. Figueira. Multiobjective Discrete and
Combinatorial Optimization. Annals of Operations Research 2005.
M. Ehrgott, X. Gandibleux. Approximative solution methods for
multiobjective combinatorial optimization. TOP 12(1):1-88, 2004.
X. Gandibleux, A. Jaszkeiwicz, A. Fréville. Multiple Objective
Metaheuristics. Journal of Heuristics 6(3), 2000.
X. Gandibleux, M. Sevaux, K. Sörensen, V. T’Kindt. Metaheuristics
for Multiobjective Optimisation. Springer 2004.
J. Figueira, S. Greco, M. Ehrgott. Multiple Criteria Decision
Analysis. Springer 2005.
X. Gandibleux. MCDM Numerical Instances Library.
www.terry.uga.edu/MCDM.

Introduction
Solution Methods
Applications
Conclusion

Multiobjectivo

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Multiobjectivo

Enviado por

Direitos autorais:

Formatos disponíveis

Introduction

Multiobjective Combinatorial Optimization

Department of Engineering Science

Operations Research 2005, Bremen, 7. – 9. September 2005

Matthias Ehrgott Multiobjective Combinatorial Optimization

with Xavier Gandibleux, Laboratoire d’Informatique de Nantes

Matthias Ehrgott Multiobjective Combinatorial Optimization

x ∈ {0, 1}n −→ n variables, i = 1, . . . , n

Matthias Ehrgott Multiobjective Combinatorial Optimization

x ∈ {0, 1}n −→ n variables, i = 1, . . . , n

Matthias Ehrgott Multiobjective Combinatorial Optimization

x ∈ {0, 1}n −→ n variables, i = 1, . . . , n

Matthias Ehrgott Multiobjective Combinatorial Optimization

feasible set in decision space 6

feasible set in objective space 4

Matthias Ehrgott Multiobjective Combinatorial Optimization

feasible set in objective space 4

Matthias Ehrgott Multiobjective Combinatorial Optimization

feasible set in decision space 6

feasible set in objective space 4

Matthias Ehrgott Multiobjective Combinatorial Optimization

zk (x̂) ≤ zk (x) for all x ∈ X 8

Lexicographic optimality (1) 6

z(x̂) ≤lex z(x) for all x ∈ X 5

Lexicographic optimality (2) 4

z π (x̂) ≤lex z π (x) for all x ∈ X 3

and some permutation z π of 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

zk (x̂) ≤ zk (x) for all x ∈ X 8

Lexicographic optimality (1) 6

z(x̂) ≤lex z(x) for all x ∈ X 5

Lexicographic optimality (2) 4

z π (x̂) ≤lex z π (x) for all x ∈ X 3

and some permutation z π of 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

zk (x̂) ≤ zk (x) for all x ∈ X 8

Lexicographic optimality (1) 6

z(x̂) ≤lex z(x) for all x ∈ X 5

Lexicographic optimality (2) 4

z π (x̂) ≤lex z π (x) for all x ∈ X 3

and some permutation z π of 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

Weakly efficient solutions XwE

There is no x with z(x) < z(x̂)

z(x̂) is weakly nondominated

There is no x with z(x) ≤ z(x̂) 3

and z(x) 6= z(x̂) 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

Weakly efficient solutions XwE

There is no x with z(x) < z(x̂)

z(x̂) is weakly nondominated

There is no x with z(x) ≤ z(x̂) 3

and z(x) 6= z(x̂) 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

Supported efficient solutions

XSE : There is λ > 0 with

face of conv(Y ) + Rp= → XSE 2 4

Nonsupported efficient solutions 2

Matthias Ehrgott Multiobjective Combinatorial Optimization

Supported efficient solutions

XSE : There is λ > 0 with

Matthias Ehrgott Multiobjective Combinatorial Optimization

Supported efficient solutions