Creating Adaptive Game AI in a Real Time Continuous Environment using Neural Networks

Abstract

This thesis gives some background on the use of Artificial Intelligence techniques in game development, notably reinforcement learning techniques implemented to allow games to learn to play and improve via selfplay, or playing against a human opponent and artificial neural networks used in the decision making process of ingame agents. Some of these ideas and techniques, originally developed for symbolic games are then applied to a real time, continuous game akin to modern, commercial video games. A modular artificial neural network architecture is used to create AI agents, capable of not only showing visibly intelligent behaviour, but also of adapting to changing game parameters in the game via online learning algorithms. The concept of controlled learning is introduced to make the use of learning AI agents more attractive to game developers.