Creative Behaviour in Evolving Agents

Abstract

Computational creativity is a field of research that is primarily focused on music composition, art and linguistics. However, research pertaining to computational creative behaviour is sparse. There is no clear, established methodology when it comes to making agents behave creatively, or how to apply machine learning techniques to make them do so. Using Evolutionary Algorithms (EAs), agents were evolved using different modular parts. The parts were several sensors and actuators that the agents used to solve tasks given to them. Their behaviour was determined by three different Artificial Intelligence (AI) methods. This thesis looks into how these three AI methods, namely Continuous Time Recurrent Neural Network (CTRNN), Fuzzy Logic, and NeuroEvolution of Augmenting Topologies (NEAT), performed when evolving behaviour in the evolving agents. In order to investigate this, a system capable of evolving and visualising agents in different environments was developed. The results of using these methods show few signs of the evolution of creative behaviour. Instead, the agents simply evolved to solve a given task. However, some unexpected actions can arguably be considered creative. This thesis contains the following contributions: A study of how different AI methods perform when evolving creatures to behave creatively, and an implemented system that evolves creatures in different environments using the three AI methods mentioned above.