Iterative Structure-Based Genetic Programming for Neural Architecture Search

In this paper we present an iterative structure-based genetic programming algorithm for neural architecture search. Canonical genetic programming uses a fitness function to determine where to move the search to in the program space. This research investigates using the structure of the syntax trees, representing different areas of the program space, in addition to the fitness function to direct the search. The structure is used to avoid areas of the search that previously led to local optima both globally (exploration) and locally (exploitation). The proposed approach is evaluated for image classification and video shorts creation. The iterative structure-based approach was found to produce better results then canonical genetic programming for both problem domains, with a slight reduction in computational cost. The approach also produced better results than genetic algorithms which are traditionally used for neural architecture search.