In this paper, a novel method of symbolic feature construction for machine learners is proposed. The method uses evolutionary computation to evolve feature transformation expressions, encoded in a form of fixed-length sequences of predefined elementary operations. Two variants of the method are proposed. One of them involves evolutionary computation for searching the space of possible solutions, whereas the other one engages cooperative coevolution for the same purpose. The results of extensive experimental evaluation on reference machine learning problems indicate superiority of the coevolutionary variety of the proposed approach.
@INPROCEEDINGS { KrawiecWlodarski04,
ABSTRACT = { In this paper, a novel method of symbolic feature construction for machine learners is proposed. The method uses evolutionary computation to evolve feature transformation expressions, encoded in a form of fixed-length sequences of predefined elementary operations. Two variants of the method are proposed. One of them involves evolutionary computation for searching the space of possible solutions, whereas the other one engages cooperative coevolution for the same purpose. The results of extensive experimental evaluation on reference machine learning problems indicate superiority of the coevolutionary variety of the proposed approach. },
AUTHOR = { K. Krawiec and L. W{\l}odarski },
BOOKTITLE = { New Trends in Intelligent Information Processing and Web Mining },
PAGES = { 139--150 },
PUBLISHER = { Springer-Verlag },
SERIES = { Advances in Soft Computing Series },
TITLE = { Coevolutionary feature construction for transformation of representation },
YEAR = { 2004 },
}