The distributed evolutionary computation platform EC-Star is extended in this paper to probabilistic classifiers. This extension, called PRETSL, allows the distributed age-layered evolution of probabilistic rule sets, which in turn makes more fine-grained decisions possible. The method is tested on 20 UCI data problems, as well as a larger dataset of arterial blood pressure waveforms. The results show consistent improvement in all cases compared to binary classification rule-sets. Probabilistic rule evolution is thus a promising approach to difficult classification tasks and particularly well suited for time-series classification.
This work was done at Sentient Technologies, Inc. and University of Oxford.