Equilibrium Finding in Large Imperfect- Information Games
Real-world strategic interactions typically involve hidden information. They can be effectively modeled as sequential imperfect-information games. We introduce several contributions to solving such games. On the abstraction side, we introduce new techniques to abstract games in a principled way without relying on human domain knowledge. This is accomplished by interleaving the abstraction step with the equilibrium finding process. These techniques were used to develop Libratus, the first AI to defeat top humans in the benchmark challenge problem of heads-up no-limit Texas hold'em poker.
Thesis Committee: Tuomas Sandholm (Chair) Ariel Procaccia Geoff Gordon Michael Wellman (University of Michigan) Satinder Singh (Uiversity of Michigan)