Integrating Feature Selection into Program Learning

Abstract

In typical practical applications of automated program learning, the scope of potential inputs to the programs being learned are narrowed down during a preliminary step. However, this approach will not work if one wishes to apply program learning as a key component of an AGI system, because there is no generally applicable feature selection heuristic, and in an AGI context one cannot assume a human cleverly tuning the feature selection heuristics to the problem at hand. A novel technique, LIFES (Learning-Integrated Feature Selection), is introduced here, to address this problem. In LIFES, one integrates feature selection into the learning process, rather than doing feature selection solely as a preliminary stage to learning. LIFES is applicable relatively broadly, and is especially appropriate for any learning problem possessing properties identified here as data focusable and focusable. It is also applicable with a wide variety of learning algorithms, but for concreteness is presented here in the context of the MOSES automated program learning algorithm. To illustrate the general effectiveness of LIFES, example results are given from applying MOSES+LIFES to gene expression classification. Application of LIFES to virtual and robotic agent control is also discussed.