Particle Swarm Stepwise (PASS) Algorithm for Variable Selection

• Activity Number: 245
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Ray-Bing Chen*+ and Chien-Chih Huang and Weichung Wang
• Companies: National Cheng Kung University and National Taiwan University and National Taiwan University
• Keywords: information criterion ; stochastic stepwise selection ; variable selection ensemble

Abstract

A new stochastic search algorithm is proposed for information criterion variable selection problems. The proposed algorithm integrates the stochastic stepwise selection approach with the particle swarm optimization, and is named as particle swarm stepwise algorithm. The key idea of the proposed algorithm is to search the best model for a pre-specified information criterion by quickly exploring the candidate model space from multiple start models and sharing the search information among all individual search paths. In addition to directly solve the information criterion variable selection problems, the proposed algorithm can also be used to generate the variable selection ensembles. Several examples are used to demonstrate the performances of our proposed algorithm.