Evolutionary development of robotic organisms

Abstract

Thirteen years have passed since Karl Sims published his work on evolving virtual creatures. Since then, several approaches to neural network evolution and genetic algorithms have been introduced. This thesis proposes a novel algorithm for the evolution of virtual creatures. The algorithm - Hierarchical NEAT - is inspired by NeuroEvolution of Augmenting Topologies (NEAT) algorithm which efficiently evolves artificial neural networks. Hierarchical NEAT applies all three main components of NEAT algorithm (protecting evolutionary innovation through speciation, sensible mating of the creatures and incremental growth from minimal structure) to the evolution of morphology and control system of the virtual creatures. Furthermore, the algorithm also allows sensible mating of control systems of the creatures, as opposed to original mating methods. Experiments have shown that the proposed algorithm significantly increases the performance of the evolution on all tested tasks. Several supplementary experiments have also been conducted to confirm that each component of the algorithm is beneficent for the evolution, that central coordination in not necessary for successful evolution of light-following strategies and that the choice of neuron transfer functions does not have significant impact on the evolution of the...