Evolving Intelligent Multimodal Gameplay Agents and Decision Makers with Neuroevolution

Abstract

�Super Mario Bros� is a difficult platforming game that requires the use of multiple behavioral modes to complete different gameplay elements such as: collecting coins, dodging enemies and getting to the end of the level. Methods for creating intelligent game playing agents have previously used human designed behavior policy for each gameplay state or by combining gameplay goals into a single task to be learned. This thesis assesses the development and method of training machines to promote multiple modes of behavior within neural network controllers. These controllers utilize the concept of evolution through multi-objective optimization for the test bench platform game system �MarioAI�. Artificial neural networks were evolved to exhibit complex and multimodal behavior using multiple sub objectives of the game; and thus overcome the non-linear, noisy, and fractured game environment. Experiments were conducted with the purpose of creating multiple Pareto-optimal solutions of quality with differing behavioral aspects. These solutions were then discerned by a Decision Maker Neural Network Ensemble that had been evolved to pick the best solution according to game level. This Decision Maker Ensemble proved to be able to learn on minimal information and provide the highest overall game score. The results of this thesis show that it�s possible to train agents on sub objectives to teach multiple forms of complex behavior that can then be abstractly chosen by an evolved Decision Maker to provide a better outcome than agents that were trained specifically towards that single solution.