Machine Learning for Games
The course gives an introduction to the application of machine learning techniques to games. The course will consist of two parts, part I dealing with computer/video games, part II dealing with traditional board/strategy games. Alongside, I will introduce necessary background material including aspects of neural networks, reinforcement learning, and graphical models. 1. In recent years various aspects of computer games have been developed to near perfection. These include high-performance graphics, realistic surround sound, and detailed physical simulations. However, the control of non-player characters (NPCs), also known as game AI, has fallen behind to the point that the resulting gaming experience often suffers. Machine learning offers a framework for making NPCs adaptive to both the environment and the human player. This technology has therefore the potential to greatly enhance gaming experience. Furthermore, at development time machine learning techniques can be employed to automate the creation of (intelligent) NPC behavior, thereby replacing the current standard of scripting and trial-and-error. The examples presented include imitation learning for avatars and reinforcement learning in fighting games. 2. Classical board games such as Chess, Go, and Backgammon have been a traditional theme in artificial intelligence. While chess has essentially been solved by traditional AI approaches, world-class Backgammon engines could only be developed based on machine learning techniques, originally in the combination of neural networks and reinforcement learning. For the traditional board game Go, neither of the two approaches has been successful so far. In this part of the course I will explain and discuss the machine learning approach to Backgammon. I will then give an introduction to the game of Go and discuss what machine learning may be able to contribute to the field of computer Go with a particular focus on modeling the uncertainty that emerges from the game's overwhelming complexity.
Author: Thore Graepel, Microsoft Research