next up previous
Next: Acknowledgements Up:

Better Trained Ants Previous: One Point Crossover


Conclusions

 

By making a modest change to the Santa Fe trail problem we have made it significantly easier for GP. In terms of the effort required (120,000) it is approximately three times easier and GP performs about as well as the best results previously obtained using variable length hill climbers (187,500) [Langdon1998c,Langdon and Poli1998] and size restricted EP search (136,000) [Chellapilla1997]. By enforcing an optimal size limit on the programs being evolved we might be able to do marginally better (104,000, size ).

Surprisingly GP performance is only marginally affected by a size limit and is roughly constant for wide ranges in maximium size.

Using the improved fitness function on the original search space, GP is able to out perform the best that might be obtained by random search by about 4:1. If both search techniques are limited to programs of a maximum length of 50 (which best suits GP) this ratio expands to about 25:1.



William B Langdon
Thu Apr 2 11:51:42 BST 1998