Created by W.Langdon from gp-bibliography.bib Revision:1.7954
This paper investigates three new extensions to Chameleon's original simple setup, seeking ways for an even more effective local search. These are: trying alternative, more cost-reflective parsimony measures such as visitation length instead of tree size; varying the reward function to more gently encourage parsimony than that in the original setup; and having more tuning earlier in runs when smaller trees can be tuned more cheaply and effectively. These strategies were tested on a varied suite of 16 difficult artificial and real-world regression problems. All of these techniques improved performance.
We show that these strategies successfully combined to cumulatively improve average test RMSE by 31percent over the original and already effective Chameleon tuning scheme. A minimum of 64 simulations were run on every problem/tuning setup combination.",
Genetic Programming entries for Fergal Lane R Muhammad Atif Azad Conor Ryan