- author = "Kyle Cranmer and R. Sean Bowman",
- title = "{PhysicsGP}: A Genetic Programming approach to event selection",
- journal = "Computer Physics Communications",
- year = "2005",
- volume = "167",
- number = "3",
- pages = "165--176",
- month = "1 " # may,
- keywords = "genetic algorithms, genetic programming, Triggering, Classification, VC dimension, Neural networks, ANN, Support vector machines, SVM",
- ISSN = "0010-4655",
-
URL = "
http://arxiv.org/abs/physics/0402030",
-
DOI = "
10.1016/j.cpc.2004.12.006",
- abstract = "We present a novel multivariate classification technique based on Genetic Programming. The technique is distinct from Genetic Algorithms and offers several advantages compared to Neural Networks and Support Vector Machines. The technique optimises a set of human-readable classifiers with respect to some user-defined performance measure. We calculate the Vapnik-Chervonenkis dimension of this class of learning machines and consider a practical example: the search for the Standard Model Higgs Boson at the LHC. The resulting classifier is very fast to evaluate, human-readable, and easily portable. The software may be downloaded at: http://cern.ch/~cranmer/PhysicsGP.html.",
-
notes = "replaces oai:arXiv.org:physics/0402030
http://www.elsevier.com/wps/find/journaldescription.cws_home/505710/description#description
p171 {"}It is meaningless to calculate the VCD (Vapnik-Chervonenkis dimension) for GP in general...{"} {"}by placing a bound on either size... or the degree of the polynomial, we can calculate a sensible VCD.{"}
GP compared with ANN (backprop + momentum) and SVM with RBF kernel (BSVM-2.0). Training data subsampled.
p174 {"}GP approach does not seem particularly sensitive to the size penalty of mutation rates{"}.",
