Genetic Programming Tree Shape Evolution

click image to reload

The particulars of the genetic programming system are: populations size of 500, generational, initialisation with ramped half-n-half (sizes between depth 2 and 4), subtree crossover with internal node selection of 0.9 and maximum size limit of depth 10 for offspring, and finally, tournament selection with size 4. Normal even-parity parameters were used.

Home

This movies illustrates the evolving tree shape of the best-of-generation solution in a run of genetic programming for the Even-5-Parity Problem. The first improving solution in each generation is selected and its shape shown in the circular lattice visualisation.

Interesting things to note:

increasingly fixed root structure over time,
occasional reduction in size in later generations,
many small changes to leaf-portions of tree.

I do not show the content at each node location, only the shape of the solution. The best fitness in this particular run improved from around 16/32 test cases correctly classified to approximately 24/32. Not a particularly great run, but representative of these system settings.

Return to Visualisations

See My Publications

Learn more about Genetic Programming