66 lines
1.9 KiB
Plaintext
66 lines
1.9 KiB
Plaintext
|
+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
|
||
|
|
|
||
|
|
Co-evolutionary Symbolic Regression (coev_symbreg):
|
||
|
|
Mixed real-valued GA - GP co-evolution example with Open BEAGLE
|
||
|
|
|
||
|
|
Copyright (C) 2003
|
||
|
|
by Jiachuan Wang <jiacwang@ecs.umass.edu>
|
||
|
|
and Christian Gagne <cgagne@gmail.com>
|
||
|
|
|
||
|
|
+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
|
||
|
|
|
||
|
|
|
||
|
|
Getting started
|
||
|
|
===============
|
||
|
|
|
||
|
|
Example is compiled in binary 'coev_symbreg'. Configuration file for
|
||
|
|
real-valued GA population (the evolving training set) is in file
|
||
|
|
'trainset-thread.conf', while configuration file for GP symbolic
|
||
|
|
regression population is in file 'symgp-thread.conf'.
|
||
|
|
|
||
|
|
Objective
|
||
|
|
=========
|
||
|
|
|
||
|
|
Two populations competitive co-evolution for symbolic regression.
|
||
|
|
First population is made of standard GP mathematical expressions.
|
||
|
|
Second population is made of individual representing 20 $(x_i,y_i)$
|
||
|
|
samples in the domain $[-1,1]$ of the equation $x^4+x^3+x^2+x$.
|
||
|
|
The global objective is to a find symbolic expressions that
|
||
|
|
"generalizes" well over the domain sampled.
|
||
|
|
|
||
|
|
Terminal set of GP expressions
|
||
|
|
==============================
|
||
|
|
|
||
|
|
X (the independent variable)
|
||
|
|
Ephemeral constants randomly generated in $[-1,1]$
|
||
|
|
|
||
|
|
Function set of GP expressions
|
||
|
|
==============================
|
||
|
|
|
||
|
|
+
|
||
|
|
-
|
||
|
|
*
|
||
|
|
/ (protected division)
|
||
|
|
|
||
|
|
Representation of training sets
|
||
|
|
===============================
|
||
|
|
|
||
|
|
Vector of 20 real-valued numbers in $[-1,1]$.
|
||
|
|
|
||
|
|
Fitness
|
||
|
|
=======
|
||
|
|
|
||
|
|
Root Mean Square (RMS) error of the GP expressions on the 20 samples
|
||
|
|
of the training set. GP expression minimizes the RMS error, while
|
||
|
|
training set maximizes it. Individuals of one population are
|
||
|
|
evaluated against last generation other population best performing
|
||
|
|
individual.
|
||
|
|
|
||
|
|
Reference
|
||
|
|
=========
|
||
|
|
|
||
|
|
Liviu Panait and Sean Luke, Methods for Evolving Robust Programs,
|
||
|
|
Proceedings of Genetic and Evolutionary Computation -- GECCO-2003,
|
||
|
|
LNCS, Vol. 2724, pp. 1740-1751, Springer-Verlag, 12-16 July 2003.
|
||
|
|
|