Teaching artificial agents to play complex video games from demonstrations

The goal of this research project is to develop novel technics to teach artificial agents how to play complex video games using reinforcement learning and demonstrations. Namely, we wish to propose a novel approach for learning from demonstrations, in which an agent simultaneously learns a behavior and the corresponding reward signal. This training procedure will rely on generative adversarial imitation learning in order to learn from expert demonstrations (in our case from players). We also want to tackle the problem of player-artificial agent cooperation, where we want to improve the robustness of current imitation learning techniques to mimic a playerÂ’s gameplay by extending the imitation learning procedure with reinforcement learning.

Intern: 
Paul Barde
Faculty Supervisor: 
Christopher Pal
Province: 
Quebec
Partner: 
Partner University: 
Program: