Adaptive sampling for active learning with genetic programming

Active learning is a machine paradigm allowing to decide which inputs use for training. It introduced Genetic Programming (GP) essentially thanks the dynamic data sampling, used address some known issues such as computational cost, over-fitting problem and imbalanced databases. The traditional sampling GP gives algorithm new sample periodically, often each generation, without considering state of evolution. In so doing, individuals do not have enough time extract hidden knowledge. An alternative approach information about adapt periodicity training change. this work, we propose an adaptive strategy classification tasks based on solved fitness cases throughout learning. flexible that could be applied with any sampling. We implemented algorithms extended controlling re-sampling frequency. experimented them solve KDD intrusion detection Adult incomes prediction problems GP. experimental study demonstrates how frequency control preserves power possible improvements in quality. also demonstrate can multi-level This work opens many relevant extension paths.