Believable Characters in Video Games
This thesis focuses on the design of believable characters in virtual envi- ronments. However, there are a lot of different environments. The choice of one environment has a lot of impact on the definition of the believability. Therefore we must choose a kind of environment to work with before going further.
Research on believable characters in virtual reality is often linked to video games (McMahan, 2003). The first reason is because game designers really need believable characters to give the best gaming experience to the players. The second reason is because scientists need a complete and easy-to-use envi- ronment where to develop believable characters which they can found in video games (Laird and Lent, 2001). Video games are thus a very useful platform to work on believability and to take inspiration from. We will now use the word player to designate a human user of a video game. As believable characters are the very goal of our work, we need character centric video games. For their behaviours to express fully, agents should be able to interact with the environment, other agents and players. The more complex and the higher the number of interactions, the harder the believability is to achieve. However, speech is a very particular form of interaction and we do not want to handle that one for now. Finally, in this thesis, we focus on open environments where the programmer cannot predict what will happen. Such games are often called sandbox games. Therefore we need a sandbox video game based on characters with lots of interactions except speech. The video game industry being very prolific, we first need to categorize video games before choosing one for our application. Laird and Lent (2001) and Mac Namee (2004) had already defined different groups:
• Action: players and agents take the role of fighters, running around and trying to kill one another.
• Role playing: players take the role of heroes in fantastic or futuristic environments. They must use diplomacy and strength to make their way to victory.
• Adventure: players follow a plot and solve puzzles and riddles.
• Strategy: players command armed units to fight battles, build bases and manage resources.
• God games: players have god-like control over the environment to build cities, shape the landscape and rule over populations.
Behaviour Models for Believable Agents
As there are very few evaluations of the believability of behaviour models, we will try to find the models which fulfil most of the re- quirements for believability. We must keep in mind that as long as
the behaviours generated by the model are not evaluated, the model cannot be considered to be a solution to our problem. However, eval- uating each model is infeasible because it would need too much time and resources.
First we will list the requirements models for believable agents must fulfil and categorize potential solutions in section 2.1.1. Then, we will analyse the four categories, connectionist models in section 2.1.2, state transition systems in section 2.1.3, production systems in section 2.1.4 and finally probabilistic models in section 2.1.5.
Requirements and Possible Solutions
Because of the context of this thesis, we will only look at models for em- bodied agents in the following study. Those models can handle interactions with virtual environments and avatars. Therefore, they all fulfil the require- ment [B1: Reaction]. According to the criteria we listed in section 1.2, we list requirements for the model itself in the table 2.1.
In these criterion, we do not take into account [B6: Perception] because we consider it is not the role of the behaviour model to model human-like perception. This criterion will be useful when we will have to choose which data will be given to the model in the implementation. As we already stated, there are a lot of available models for the control of behaviour. Instead of examining each model, we will rather study models by category. We choose to categorize behaviour models into 4 types:
• Connectionist models.
• State transition systems.
• Production systems.
• Probabilistic models.
Probabilistic models make use of random variables and discrete or contin- uous probability distributions to generate output values. The distributions form the parameters of the model, defining the relation between each random variable (see figure 2.5). They have been very used, for a long time now, in hand writing (Bozinovic and Srihari, 1982; Vinciarelli, 2002; Artieres et al., 2007) and speech recognition (Levinson, 1983; Mari et al., 1996; Glass et al., 1996). Later, their use spread to robotics for the control of movement (Simmons and Koenig, 1995; Calinon and Billard, 2007) and then to the control the whole behaviour of real or virtual agents (Le Hy et al., 2004; Gorman et al., 2006a; Bauckhage et al., 2007). In order to control the behaviour of agents, probabilistic models have to answer to the question P(Actions|Sensors). It can be translated as “what is the probability of doing actions given the current value of the sensors”. Note that is may be seen as an extension of a rule IF Sensors THEN Actions, because the probability can be translated as IF Sensors THEN you may do Actions. Of course, models can have internal random variables, the question being P(Actions|Sensors, InternalState).
Table of contents :
List of Figures
List of Tables
List of Acronyms
List of Notations
1.1 From Virtual Reality to Believable Agents in Video Games
1.2 Assessing the Believability of Characters in Video Games
1.3 Objective of our Work
1.4 Organization of this Manuscript
2 Behaviour Models and Learning Algorithms for Believable Agents
2.1 Behaviour Models for Believable Agents
2.2 Algorithms to Learn Behaviours
2.3 Le Hy’s Work
2.4 Believability of Agents Using Le Hy’s Model
3 Chameleon: Behaviour Model and Learning Algorithm for Believable Agents
3.1 Semantic Refinement
3.2 Attention Selection Mechanism
3.3 Learning the Environment
3.4 Learning the Model Parameters via an EM Algorithm .
4 Analysis and Evaluation of an Implementation of Chameleon
4.1 Semantic Refinement
4.2 Attention Selection Mechanism
4.3 Learning The Environment
4.4 Learning the Parameters of the Model with an EM Algorithm .
5.4 Future Work