Path length, or search complexity, is an understudied property of trees in genetic programming. Unlike size and depth measures, path length directly measures the balancedness or skewedness of a tree. Here a close relative to path length, called visitation length, is studied. It is shown that a population undergoing standard crossover will introduce a crossover bias in the visitation length. This bias is due to inserting variable length subtrees at various levels of the tree. The crossover bias takes the form of a covariance between the sizes and levels in the trees that form a population. It is conjectured that the crossover bias directly determines the size distribution of trees in genetic programming. Theorems are presented for the one-generation evolution of visitation length both with and without selection. The connection between path length and visitation length is made explicit.
[1]
Riccardo Poli,et al.
The evolution of size and shape
,
1999
.
[2]
George R. Price,et al.
Selection and Covariance
,
1970,
Nature.
[3]
Lee Altenberg,et al.
The Schema Theorem and Price's Theorem
,
1994,
FOGA.
[4]
J. K. Kinnear,et al.
Advances in Genetic Programming
,
1994
.
[5]
Terence Soule,et al.
Genetic Programming: Theory and Practice
,
2003
.
[6]
Riccardo Poli,et al.
A Schema Theory Analysis of the Evolution of Size in Genetic Programming with Linear Representations
,
2001,
EuroGP.
[7]
L. Altenberg.
The evolution of evolvability in genetic programming
,
1994
.
[8]
Mark Kotanchek,et al.
Pareto-Front Exploitation in Symbolic Regression
,
2005
.
[9]
Philippe Flajolet,et al.
An introduction to the analysis of algorithms
,
1995
.