History of Articial Intelligence in Video Games

Get Complete Project Material File(s) Now! »

Imitation Learning: The Next Challenge of  Video Games’ AI?

This review allows us to identify the main uses of AI in video games as well as the constraints on the techniques to be used. As a summary of that review, we have seen that AI in video games aims to:
Improve the performance of non-player characters (as opponents or friends).
Enhance the believability of NPCs.
Propose innovative gameplay.
Promote replay value.
Adjust the game’s diculty.
Automatically generate content (maps, story arc, etc)
Solve game-specic problems (pathnding, choosing an optimal strategy, etc) If the answers to game-specic problems and automatic content generation seem to be side issues, the other use cases may be addressed by a generic imitation learning solution. Indeed, with an eective solution, the performance of the reproduced behavior would be directly related to the level of the teacher player. In the case of online learning, it could be possible to propose gameplay based on this learning (as in Creatures, Black & White or Forza Motors). Moreover, online learning with the player as the teacher would cause a change in the NPCs’ behavior. Therefore the player would still face a NPC with adapted skills. Still in the case of oine learning, real life data could be used. It would then be possible to learn the behavior of an entire football team or of a real tennis player for example. The development of a generic software suite for learning behavior by imitation would address, to some extent, these questions. The development of an articial behavior would then amount to a learning session. One may compare this mode of development to motion capture. This process has resulted in signicant time saving in the production of humanoid animations. Without motion capture, a graphic designer has to decompose all movements manually. Motion capture can signicantly speed up the animation production process and allow even more believable animations. Behavior capture seems to be an interesting option for the future of AI in computer games. To the best of our knowledge, the result of current imitation learning techniques are far from convincing in all situations. The aim of this thesis is to propose ways to make a further step towards this challenge.

A Modern AI System for Video Games: Requirements

The design of a generic Articial Intelligence (AI) solution for video games needs to take many requirements into account. First (section 2.1.1), we list the development constraints and identify a paradigm that would suit our needs. Then (section 2.1.2), as we want to develop a learning system, we dene criteria related to the characteristics of the behaviors we want to reproduce. Finally (section 2.1.3), we consider learning problems. These two types of requirements will allow us to identify the strengths and weaknesses of the dierent methods we may use.

READ  Features of ultrasonic wave phase conjugation system operating with weak signals

Development Requirements

As in any complex project, the development of a video game involves many people over a long period of time. Many constraints appear relating to the general architecture of the solution. The associated requirements that we have identied are listed in table 2.1.

General Models for AI Systems

We have demonstrated, in the previous chapter, the value of using data mining to build our solution. Classication and regression algorithms can reproduce reactive behaviors ([E1: Reaction]) and some of them are able to partly produce variability ([E2: Variability]). The extraction of knowledge can be implemented through vector quantization techniques while providing [E4: Memory] to the solution. When these techniques are used [L1: Online], the learned behavior can get better in time ([E6: Evolution]). However, these techniques are not, strictly speaking, a technical development that meets the requirements dened in section 2.1.1. In addition, some expressivity requirements dened in section 2.1.2 are not met. [E5: Planning], [E7: Inter Specic Variability], [E8: Organization] or [E9: Coordination] are indeed dicult to learn automatically. Because we want to build a solution able to learn this expressivity, we must implement the use of data mining solution in a more general architecture, responding to development criteria such as [D1: Maintainability], [D3: Flexibility] or [D5: Communication Support]. In this section, we study dierent families of architectures that can be implemented as part of an AI solution in video games. This study will allow us to choose the most suitable architecture to our requirements.

Table of contents :

List of Figures
List of Tables
List of Acronyms
1 Introduction 
1.1 History of Articial Intelligence in Video Games
1.2 Imitation Learning: The Next Challenge of Video Games’ AI? .
1.3 Objective of our Work
1.4 Organization of this Manuscript
2 AI Solutions for Video Games 
2.1 A Modern AI System for Video Games: Requirements
2.2 Data Mining
2.3 Discussion
3 Orion: A Generic Model Proposition for Imitation Learning of Behavior 
3.1 General Models for AI Systems
3.2 Orion Structural Model
3.3 Orion Behavioral Model
3.4 Conclusion
4 Applications: Pong and Unreal Tournament 3 
4.1 Pong
4.2 Unreal Tournament 3
4.3 Conclusion
5 Conclusion 
5.1 Summary of the Thesis
5.2 Discussion
5.3 Future Work
5.4 Publications


Related Posts