Bayesian Optimization Algorithms for Multi-Objective Optimization

journal article - other

English

In recent years, several researchers have concentrated on using probabilistic models in evolutionary algorithms. These Estimation Distribution Algorithms (EDA) incorporate methods for automated learning of correlations between variables of the encoded solutions. The process of sampling new individuals from a probabilistic model respects these mutual dependencies among genes such that disruption of important building blocks is avoided, in comparison with classical recombination operators. The goal of this paper is to investigate the usefulness of this concept in multi-objective evolutionary optimization, where the aim is to approximate the set of Pareto-optimal solutions. We integrate the model building and sampling techniques of a special EDA called Bayesian Optimization Algorithm based on binary decision trees into a general evolutionary multi-objective optimizer. A potential performance gain is empirically tested in comparison with other state-of-the-art multi-objective EA on the bi-objective 0/1 knapsack problem.

probabilistic models,Estimation Distribution Algorithms, multi-objective evolutionary optimization, Pareto-optimal solutions, Bayesian Optimization Algorithm, binary decision trees, knapsack problem.

LAUMANNS, M., OČENÁŠEK, J.

2004

7. 9. 2002

Springer Verlag

Granada

3-540-444139-5

Parallel Problem Solving from Nature - PPSN VII

0302-9743

Lecture Notes in Computer Science

2002

2439

Federal Republic of Germany

298

307

10