Active learning for genetic programming using model ensemble uncertainty was explored across a range of uncertainty metrics to determine if active learning can be used with GP to minimize training set sizes by selecting maximally informative samples to guide evolution. The choice of uncertainty metric was found to have a significant impact on the success of active learning to inform model development in genetic programming. Differential evolution was found to be an effective optimizer, likely due to the non-convex nature of the uncertainty space, while differential entropy was found to be an effective uncertainty metric. Uncertainty-based active learning was compared to two random sampling methods and the results show that active learning successfully identified informative samples and can be used with GP to reduce required training set sizes to arrive at a solution.
In this contribution we study how to effectively evolve programs tailored for biomedical image segmentation by using an Active Learning approach in Cartesian Genetic Programming (CGP). Active Learning allows to dynamically select training data by identifying the most informative next image to add to the training set. We study how different metrics for selecting images under active learning impact the searchability of CGP. Our results show that datasets built during evolution with active learning improve the performance of Cartesian GP substantially. In addition, we found that the choice of the particular metric used for selecting which images to add heavily impacts convergence speed. Our work shows that the right choice of the image selection metric positively impacts the effectiveness of the evolutionary algorithm.