Robot Soccer With GNG & NEAT

Complexification is an essential concept in evolutionary robotics that develops dynamic neural networks over generations with the aim of increasing their capabilities. Artificial networks networks are complexified by altering, adding and removing nodes and connections from the neural network structure. The dimensionality of the input layer in a neural network also plays a significant role in the development of intelligent behavior in robots. Reducing dimensionality is proven to increase the speed of evolution by categorizing the robots environment. Our aim is to evolve a robot that can perform basic soccer skills by reducing the dimensionality of its input vector through categorization of its environment, and gradually complexifying the neural network using a fitness function. In phase one of our experiment, we use a low-dimensional Growing Neural Gas (GNG) to categorize the robots environment into nodes that represent highdimensional sensory inputs. In phase two, NeuroEvolution of Augmenting Topologies (NEAT) is used to evolve an intelligent brain by complexifying the neural network over evolutionary generations based on a specified fitness function. Our results showed significant increases in average and best fitness over multiple generations of evolution and supported the viability of a combination GNG and NEAT for evolving intelligent robots.