Evolving Graphs with Cartesian Genetic Programming with Lexicase Selection

The automatic construction of an image filter is a difficult task for which many recent machine-learning methods have been proposed. Cartesian Genetic Programming (CGP) has been effectively used in image-processing tasks by evolving programs with a function set specialized for computer vision. Although standard CGP is able to construct understandable image filter programs, we hypothesize that explicitly using a mechanism to control the size of the generated filter programs would help reduce the size of the final solution while keeping comparable efficacy on a given task. It is indeed central to keep the graph size as contained as possible as it improves our ability to understand them and explain their inner functioning. In this work, we use the Lexicase selection as the mechanism to control the size of the programs during the evolutionary process, by allowing CGP to evolve solutions based on performance and on the size of such solutions. We extend Kartezio, a Cartesian Genetic Programming for computer vision tasks, to generate our programs. We found in our preliminary experiment that CGP with Lexicase selection is able to achieve similar performance to the standard CGP while keeping the size of the solutions smaller.