Genetic_Programming_Theory_and_Practice_XIII

228 S. Silva et al.

Specialized class transformations Evaluated Tree

Overall Accuracy

This is equivelent to

57% M3GP

46%

61%

68% Final

Ensemble

Only if fitness

improves the

change is keept

Fig. 3 Ensemble construction to build the individual

transformationk^0 iinS, i.e.,ei Dk^0 i. After each replacement, we compute the
accuracy of the ensembleE. If the accuracy improves then the change is kept,
otherwise it is reversed. The process is depicted in Fig. 3.
There are several comments to be made regarding the proposed algorithm. First,
specialized class transformations are chosen based on the performance achieved on
each class, attempting to find improvements in terms of both TP and FP. These
criteria provide a robust estimate of performance on a class by class basis, however
it is possible that in the end we do not have the best possible transformation for
each class, but only a non-dominated individual of a larger Pareto set. Nonetheless,
we feel this selection process provides a useful first approximation. Second, we
can say that the proposed ensembles are used to construct improved versions of
each individual. This should give low quality individuals a chance to improve, and
possibly save any useful genetic material they may have. Finally, the ensemble
construction process is a greedy algorithm that may not be considering higher order
epistatic effects. Again, for now we choose the simplest approach, and will leave
future improvements as possible future research.

6 Experimental Setup

In this section we describe the data sets used for testing the methods, as well as the
tools and parameters adopted for performing the experiments.