Seminar on Computational Learning and Adaptation

 Two Dimensions for Learning
in Computer Game AI

Vincent Corruble
Laboratoire d'Informatique de Paris 6
University Pierre et Marie Curie

Abstract:

From the point of view of AI and Machine Learning, modern Computer Games constitute ideal environments for experimenting on new techniques and algorithms : they propose challenging problems for which simulation is relatively easy to setup. In this talk, I will focus on two projects which address two important dimensions of learning in games. The first one uses learning in a traditional manner to improve the performances of the computer player in a strategy game. Because of the immersing nature of the game, it requires to tackle important issues beyond the learning algorithm itself, including perceptions and automatic change of representation, as well as the issue of coordination and learning in a multi-agent setting. The other project, in collaboration with colleagues from Brazil, addresses the issue of learning not simply to optimize performances of the computer player, but rather to adapt online its performance to the level and skills of the human player (the process known as Dynamic Game Balancing).

Return to the seminar schedule