Spatial Genetic Programming

Abstract

An essential characteristic of brains in intelligent organisms is their spatial organization, in which different parts of the brain are responsible for solving different classes of problems. Inspired by this concept, we introduce Spatial Genetic Programming (SGP) - a new GP paradigm in which Linear Genetic Programming (LGP) programs, represented as graph nodes, are spread in a 2D space. Each individual model is represented as a graph and the execution order of these programs is determined by the network of interactions between them. SGP considers space as a first-order effect to optimize which aids with determining the suitable order of execution of LGP programs to solve given problems and causes spatial dynamics to appear in the system. RetCons are internal SGP operators which enhance the evolution of conditional pathways in SGP model structures. To demonstrate the effectiveness of SGP, we have compared its performance and internal dynamics with LGP and TreeGP for a diverse range of problems, most of which require decision making. Our results indicate that SGP, due to its unique spatial organization, outperforms the other methods and solves a wide range of problems. We also carry out an analysis of the spatial properties of SGP individuals.