Strategic Reparameterization for Enhanced Inference in Imperfect Information Games: A Neural Network Approach

How to infer useful information for making decisions is one of the core problems in imperfect information games. This paper tackles the problem by proposing a reparameterization technique to improve inference in imperfect information games. It constructs a continuous probability distribution that implicitly encodes both the reasoning process and the information gained during game play. This integrated distribution facilitates training neural networks for these games by capturing the inherent uncertainties. The experimental results show that this method not only improves the capability to reason about agents’ information but also could be generalized to various types of games.