Expert Iteration for Risk

Lucas Gnecco Heredia, Tristan Cazenave

Abstract: Risk is a complex strategy game that may be easier to understand for humans than chess but harder to deal with for computers. The main reasons are the stochastic nature of battles and the different decisions that must be coordinated within turns. Our goal is to create an artificial intelligence able to play the game without human knowledge using the Expert Iteration [1] framework. We use graph neural networks [13, 15, 22, 30] to learn the policies for the different decisions and the value estimation. Experiments on a synthetic board show that with this framework the model can rapidly learn a good country drafting policy, while the main game phases remain a challenge.