Comparison of a compiling genetic programming system versus a connectionist approach

Created by W.Langdon from gp-bibliography.bib Revision:1.4202

  author =       "Peter Nordin",
  title =        "Comparison of a compiling genetic programming system
                 versus a connectionist approach",
  booktitle =    "Handbook of Evolutionary Computation",
  publisher =    "Oxford University Press",
  publisher_2 =  "Institute of Physics Publishing",
  year =         "1997",
  editor =       "Thomas Baeck and David B. Fogel and 
                 Zbigniew Michalewicz",
  chapter =      "section G1.5",
  keywords =     "genetic algorithms, genetic programming, CGPS",
  ISBN =         "0-7503-0392-1",
  URL =          "",
  DOI =          "doi:10.1201/9781420050387.ptg",
  size =         "7 pages",
  abstract =     "This case study briefly presents the compiling genetic
                 programming method and evaluates its performance
                 against a neural network. Most genetic programming
                 approaches use a technique where a problem specific
                 language is executed by an interpreter. The individual
                 code segments in the population are decoded at run time
                 by a virtual machine. The disadvantage of this paradigm
                 is that interpreting the program involves a large
                 overhead. We have evaluated the idea of using the
                 lowest-level native binary machine code as the
                 individuals in the population. There is no intermediate
                 language nor any interpreting steps. The genetic
                 program that administers these machine code segments is
                 written in C. The algorithm is steady state and uses a
                 small tournament for selection. This approach has
                 enhanced performance by up to 2000 times compared to a
                 conventional system in an interpreting language. The
                 increased performance is tested on a problem of
                 symbolic regression of a classifier function in machine
                 code. We evolve a machine code program that classifies
                 Swedish words into nouns and non-nouns by spelling
                 only. We compare the compiling genetic programming
                 system (CGPS) with a neural network performing the same
                 task. In our example, the results show superior
                 performance of the CGPS compared to the connectionist
                 approach. While the classification and generalization
                 capabilities are equal, the training time is more than
                 200 times faster, the classification time 500 times
                 faster, and the memory requirements at least ten times
                 lower with the CGPS, as compared with the neural
  notes =        "training times is 200 times faster, the classification
                 times 500 times faster

                 2100 Swedish words nouns v non-nouns. 2000x faster than
                 lisp, 100x than C interpreter. Sun microsystems machine

Genetic Programming entries for Peter Nordin