AlphaJass

Abstract

Although Jass parades as a Swiss card game of national importance, it is in fact a serious problem in game theory, similar to the wire-tap problem. Jass in someways is harder to solve for perfect play than chess or Go, not because of its branching factor, which is relatively small, but because it is a hidden information game where the model has to simulate a complex posterior distribution of 280 billion possible worlds. We are solving this problem using a novel Duel Critique methodology and a multi-phase training scheme involving imitation learning, self-play, and SPARTA, that will eventually distil a super human bot team capable of beating Switzerland’s elite players.