Over the last decade, the video game industry has witnessed advancements in computer graphics, sound production and the quantitiy of available content. However, the development of agent behaviour in video games still relies on traditional methods that have been present in the industry since its very inception. To sustain the industry's growth and meet end-user satisfaction requirements in the future, innovations must also be introduced in other areas within the industry. The next generation of video games will no longer be marked by higher texture resolutions, but by realistic autonomous behaviour of agents, capable of learning and adapting to given situations. In this master's thesis we present a new hybrid method for developing believable agent behaviour in video games and compare it with existing approaches. We based the method on behaviour trees, finite state machines, and the reinforcement learning method Q-learning. The presented hybrid method combines the advantages of the listed approaches while addressing some of their weaknesses. We evaluate the behaviour believability of the new method using a questionnaire, and compare its efficiency and suitability with established approaches to developing intelligent agent behaviour.
|
