Bagging, Boosting, and Bloating in Genetic Programming

Created by W.Langdon from gp-bibliography.bib Revision:1.8051

@InProceedings{iba:1999:BBBGP,

• author = "Hitoshi Iba",
• title = "Bagging, Boosting, and Bloating in Genetic Programming",
• booktitle = "Proceedings of the Genetic and Evolutionary Computation Conference",
• year = "1999",
• editor = "Wolfgang Banzhaf and Jason Daida and Agoston E. Eiben and Max H. Garzon and Vasant Honavar and Mark Jakiela and Robert E. Smith",
• volume = "2",
• pages = "1053--1060",
• address = "Orlando, Florida, USA",
• publisher_address = "San Francisco, CA 94104, USA",
• month = "13-17 " # jul,
• publisher = "Morgan Kaufmann",
• keywords = "genetic algorithms, genetic programming, classifier ensembles",
• ISBN = "1-55860-611-4",
• URL = "http://gpbib.cs.ucl.ac.uk/gecco1999/GP-407.pdf",
• URL = "http://gpbib.cs.ucl.ac.uk/gecco1999/GP-407.ps",
• size = "8 pages",
• abstract = "We present an extension of GP (Genetic Programming) by means of resampling techniques, i.e., Bagging and Boosting. These methods both manipulate the training data in order to improve the learning algorithm. In theory they can significantly reduce the error of any weak learning algorithm by repeatedly running it. This paper extends GP by dividing a whole population into a set of sub-populations, each of which is evolvable by using the Bagging and Boosting methods. The effectiveness of our approach is shown by experiments. The performance is discussed by the comparison with the traditional GP in view of the bloating effect.",
• notes = "GECCO-99 A joint meeting of the eighth international conference on genetic algorithms (ICGA-99) and the fourth annual genetic programming conference (GP-99)

  10 Subpopulations each has its own training data (produced using the boosting or bagging methods. Best of each subpopulation has vote in final result. Do we actually need subpopulations, could not the whole algorithm be split into T entirely separate GP runs? SGPC1.1

  p1054 {"}controlling the bloating effect is closely related to the performance improvement...{"}

  noisy cos(2x)=1-sin(x)**2, Mackey-Glass chaotic time series, 6MUX, symbolic regression, nikkei225 Description of boosting weight adjustment algorithm (p1054) seems to be wrong?

  p1056 BagGP, BoostGP > GP, BagGP=BoostGP But only in the case of noisy cos(2x) does difference (table 2) seem big. Mention of DSS and PADO.

  p1059 Says Bagging and Boosting yield lower bloat. (does not explain why) Little supporting data (Fig 5). Boosting v co-evolution",

}

Genetic Programming entries for Hitoshi Iba

Citations