Neural Network Guided Evolution of L-system Plants

Abstract

A Lindenmayer system is a parallel rewriting system that generates graphic shapes using several rules. Genetic programming (GP) is an evolutionary algorithm that evolves expressions. A convolutional neural network(CNN) is a type of neural network which is useful for image recognition and classification. The goal of this thesis will be to generate different styles of L-system based 2D images of trees from scratch using genetic programming. The system will use a convolutional neural network to evaluate the trees and produce a fitness value for genetic programming. Different architectures of CNN are explored. We analyze the performance of the system and show the capabilities of the combination of CNN and GP. We show that a variety of interesting tree images can be automatically evolved. We also found that the success of the system highly depends on CNN training, as well as the form of the GP's L-system language representation.