GECCO 2004 LATE BREAKING PAPER


Repeated Sequences in Linear GP Genomes

W. B. Langdon and W. Banzhaf


Computer Science,
Memorial University of Newfoundland, Canada


W.Langdon@cs.uc1.ac.uk, banzhaf@cs.mun.ca

50% of your DNA is Repeated Sequences

How much of your programs is Repeated Sequences?

Discrete Mackey-Glass chaotic time series

The three (of twenty) amino acids plotted where selected by Discipulus as being the best discriminators. The non-linear function of number of Valines was also suggested by an evolved model.

Evolution of mean program size in ten Mackey-Glass prediction runs using two point crossover.

Evolution of length of longest repeated sequence of instructions in the best Mackey-Glass prediction program produced by the first run with two-point crossover (2XO) and fitness selection. The length of the programs is also shown.

Length of longest repeated sequence of instructions in the best prediction programs. $2\times 10$ Discipulus protien locations runs and $2\times 10$ Mackey-Glass prediction runs. All bloating runs evolve repeated sequences.

Location of repeated instructions in the best Mackey-Glass prediction program. Instructions that are part of repeated sequences longer than 10 are plotted in blue. Notice almost every instruction is repeated, so the diagonal is almost solid.

Distribution of repeated sequences along length of best Mackey-Glass predictor at end of first 2XO run. The solid line highlights the location of its 63 effective instructions. An animation can be found via http://www.cs.ucl.ac.uk/staff/W.Langdon/gecco2004lb/

Evolution of instructions in first 2XO Mackey-Glass run. The red lines and text highlight the effective instructions (e.g. a=e-(29);) in the best of run program.