| This research provides a method to enhance accuracy and reduce performance fluctuation of programs produced by genetic programming by combining individual evolved programs into robust ensembles. More effective ensembles have fewer correlated faulty outputs. Therefore, current ensemble techniques focus on diversity pressures to reduce correlated faults among the ensemble members. However, whether or not an optimal ensemble is formed through these pressures is unknown, simply because ensemble optimality is undefined.; We define the behavioral diversity of an ensemble of imperfect programs as the degree to which the ensemble failure rate deviates from what one would expect if fault occurrences were statistically independent. Given this metric, we form an ensemble by selecting individuals that exhibit this diversity from a large pool of evolved programs and combining their outputs into a single ensemble output.; Classification or prediction problems benefit the most from this research. We have validated our approach by showing statistically significant improvements when applied to a DNA segment classification problem. |
