Wednesday, January 26, 2005
Revisiting Pittsburgh
Yesterday, I was reading again Jaume's paper on incremental learning. A simple idea such as a round robin of small disjoint training sets greatly helps Pittsburgh classifier systems in the quest for generality. On the Michigan approach, however, such issue was settled down after the idea of evolving classifiers based on accuracy by Wilson (now celebrating the 10th anniversary of its publication).
The more I think about it, the more I wonder how Wilson's ideas, that gave birth to XCS, could be carried over the Pittsburgh realm, where most of the work done has focused only on evolving rule sets based on the overall performance (pure classification accuracy). If such ideas could be carried over the Pittsburgh realm, the basis for a renewed genetics-based machine learning paradigm would be sketched.
