a comparison of estimation of distribution algorithms for the linear ordering problem josu ceberio...

38
A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español de Metaheurísticas, Algoritmos Evolutivos y Bioinspirados - MAEB20

Upload: lesley-underwood

Post on 21-Dec-2015

221 views

Category:

Documents


6 download

TRANSCRIPT

Page 1: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

A comparison of estimation of distribution algorithms for the linear ordering problem

Josu Ceberio Alexander Mendiburu

Jose A. Lozano

X Congreso Español de Metaheurísticas, Algoritmos Evolutivos y Bioinspirados - MAEB2015

Page 2: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

2

Outline

• The linear ordering problem

• The Mallows and Plackett-Luce EDAs

• Experimentation

• On the Boltzmann distribution associated to the LOP

• Conclusions and future work

Page 3: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

3

Combinatorial optimization problems

Permutation optimization problemsDefinition

Page 4: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

4

Permutation optimization problemsDefinition

Problems whose solutions are naturally represented as permutations

Page 5: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

5

Permutation optimization problemsGoal

To find the permutation solution that minimizes a fitness function

The search space consists of solutions.

Page 6: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

6

Permutation optimization problemsExamples

• Travelling salesman problem (TSP)

• Permutation Flowshop Scheduling Problem (PFSP)

• Linear Ordering Problem (LOP)

• Quadratic Assignment Problem (QAP)

Page 7: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

7

Permutation optimization problemsExamples

• Travelling salesman problem (TSP)

• Permutation Flowshop Scheduling Problem (PFSP)

• Linear Ordering Problem (LOP)

• Quadratic Assignment Problem (QAP)

Page 8: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

8

The linear ordering problemDefinition

Example extracted from R. Martí and G. Reinelt (2011) The linear ordering problem: exact and heuristic methods in combinatorial optimization.

Page 9: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

9

The linear ordering problemDefinition

Example extracted from R. Martí and G. Reinelt (2011) The linear ordering problem: exact and heuristic methods in combinatorial optimization.

Page 10: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

10

The linear ordering problemDefinition

Example extracted from R. Martí and G. Reinelt (2011) The linear ordering problem: exact and heuristic methods in combinatorial optimization.

Page 11: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

11

The linear ordering problemSome applications

- Aggregation of individual preferences- Kemeny ranking problem

- Triangulation of input-output tables of the branches of an economy

- Ranking in sports tournaments

- Optimal weighted ancestry relationships

Page 12: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

12

The linear ordering problem

It is an NP-hard problem(Garey and Johnson 1979)

Page 13: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

13

Estimation of distribution algorithms Definition

Page 14: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

14

In previous works

• Implement probability models for permutation domains

– The Mallows model

– The Generalized Mallows model

– The Plackett-Luce model

Page 15: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

15

In previous works

• Implement probability models for permutation domains

– The Mallows model

– The Generalized Mallows model

– The Plackett-Luce model

Promising performanceon the LOP

Page 16: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

16

The Mallows modelDefinition

• A distance-based exponential probability model

• Central permutation

• Spread parameter

• A distance on permutations

Page 17: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

17

The Mallows modelDefinition

• A distance-based exponential probability model

• Central permutation

• Spread parameter

• A distance on permutations

Page 18: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

18

The Mallows modelDefinition

• A distance-based exponential probability model

• Central permutation

• Spread parameter

• A distance on permutations

Page 19: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

19

The Ulam distanceDefinition

Calculates the minimum number of insert operations to convert in .

Page 20: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

20

Distances and neighborhoods

– Two solutions and are neighbors if the Kendall’s-τ distance

between and is

– Two solutions and are neighbors if the Cayley distance

between and is

– Two solutions and are neighbors if the Ulam distance between

and is

Swap neighborhood

Interchange neighborhood

Insert neighborhood

Page 21: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

21

Distances and neighborhoods

– Two solutions and are neighbors if the Kendall’s-τ distance

between and is

– Two solutions and are neighbors if the Cayley distance

between and is

– Two solutions and are neighbors if the Ulam distance between

and is

Swap neighborhood

Interchange neighborhood

Insert neighborhood

Page 22: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

22

The Plackett- Luce modelDefinition

The probability of under the Plackett-Luce model is given by

The vector of scores defines the preference of each item to be ranked in

top rank

Page 23: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

23

The Plackett- Luce modelVase model interpretation

A vase of infinite colored balls

With known proportions of each color

Draw balls from the vase until a permutationof colored balls is obtained

Page 24: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

24

The Plackett- Luce modelVase model interpretation

Stage 1

We draw a ball.

The probability to extract a red ball atthis stage is:

And it is red.

Page 25: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

25

The Plackett- Luce modelVase model interpretation

Stage 2

We draw another ball.

The probability to extract a green ball from the remaining balls is:

And it is green.

Page 26: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

26

The Plackett- Luce modelVase model interpretation

Stage 3

We draw the blue ball.

The probability to extract a blue ball is:

Page 27: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

27

L-decomposability

Page 28: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

28

L-decomposability

Page 29: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

29

• Algorithms:

- Mallows EDA under the Ulam distance (MaEDA)

- Plackett-Luce EDA (PLEDA)

• 50 instances of sizes: {10, 20, 30, 40, 50, 60, 70, 80, 90, 100}

• Average Relative Percentage Deviation (ARPD) of 20 repetitions

• Stopping criterion: 100n-1 generations

ExperimentsDesign

Page 30: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

30

ExperimentsResults

Page 31: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

31

Discussion

Which is the most efficient model to optimize the LOP ?

Page 32: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

32

Discussion

Theoretically, the Boltzmann distribution associated to the LOP

Boltzmann constant

Page 33: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

33

Discussion

Calculate from the Boltzmann distribution associated to the LOP:

• the Mallows model under the Ulam distance

•the Plackett-Luce model

4 instances of size n=7

Boltzmann constant c: [0,300]

Kullback-Leibler divergence:

Learn from a sample of 106 permutations

Perform a weighted computationof the parameters

Page 34: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

34

DiscussionProbability concentrates in the fittest solutions

Near uniform distribution

Page 35: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

35

Conclusions

• For small instances, MaEDA and PLEDA obtain similar results.

• For large instances, MaEDA is the preferred algorithm.

• With respect to the Boltzmann distribution of the LOP:

– When the fitness of the solutions is very different, the Mallows model under the Ulam distance is the preferred option.

– When the fitness of the solutions is similar, the Plackett-Luce is more accurate.

Page 36: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

36

Future work

Compare Mallows EDA under the Ulam distance with state-of-the-art algorithms

Page 37: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

37

Future work

Study the properties of the Boltzmann distribution on the LOP

Page 38: A comparison of estimation of distribution algorithms for the linear ordering problem Josu Ceberio Alexander Mendiburu Jose A. Lozano X Congreso Español

A comparison of estimation of distribution algorithms for the linear ordering problem

Josu Ceberio Alexander Mendiburu

Jose A. Lozano

X Congreso Español de Metaheurísticas, Algoritmos Evolutivos y Bioinspirados - MAEB2015