Probabilistic Behaviour Model and Imitation Learning Algorithm for Believable Characters in Video Games

This manuscript aims at designing a behaviour model for the control of believable characters in video games. We define a believable character as a computer program able to control a virtual body in a virtual environment so that other human users in the environment think the virtual body is controlled by another human user. To be more precise, we define 10 requirements for a character to be believable, based on previous experiments and work. In order to fulfil these requirements, we studied the behaviour models developed both in the research and the industry. As one of the requirements is that the model is able to evolve, we had to find learning algorithms for the behaviour model. We find out that imitation is the best way to believability. With these studies in mind we find out that the behaviour model developed by Le Hy in his thesis answers to most of the requirements but has still some limitations. In this manuscript we use an approach like Le Hy's. We first try to reduce the number of parameters in the model. Then we replace the two mechanisms to break the complexity of the probability distributions by an attention selection mechanism. We add to the model the ability to learn by imitation the layout of environments. Finally we totally revamp the learning algorithm. The proposition makes the model able to learn how to act in the environment rapidly. Stimulus-action associations are made which the agent look-like a human player. However the learning also learns wrong associations which destroy the illusion of believability. According to our studies, our model performs better than Le Hy's but work has still to be done on the model to achieve the final goal.