Fitness approximation in estimation of distribution algorithms for feature selection

  • Authors:

  • Haixia Chen;Senmiao Yuan;Kai Jiang

  • Affiliations:

  • College of Computer Science and Technology, Jilin University, Changchun, China;College of Computer Science and Technology, Jilin University, Changchun, China;The 45th Research Institute of CETC, Beijing, China

  • Venue:
  • AI'05 Proceedings of the 18th Australian Joint conference on Advances in Artificial Intelligence
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

Estimation of distribution algorithms (EDAs) are popular and robust algorithms that combine two technical disciplines of soft computing methodologies, probabilistic reasoning and evolutionary computing, for optimization problems. Several algorithms have already been proposed by different authors. However, these algorithms may require huge computation power, which is seldom considered in those applications. This paper introduces a “fast estimation of distribution algorithm” (FEDA) for feature selection that does not evaluate all new individuals by actual fitness function, thus reducing the computational cost and improve the performance. Bayesian networks are used to model the probabilistic distribution and generate new individuals in the optimization process. Moreover, fitness value is assigned to each new individual using the extended Bayesian network as an approximate model to fitness function. Implementation issues such as individual control strategy, model management are addressed. Promising results are achieved in experiments on 5 UCI datasets. The results indicate that, as population-sizing requirements for building appropriate models of promising solutions lead to good fitness estimates, more compact feature subsets that give more accurate result can be found.