Information Entropy of Uncertainty Control: An Uncertainty Management Method in Imperfect Information Games

With the vigorous development of artificial intelligence, many amazing breakthroughs have been made in recent years, while games often serving as important milestones. After significant progresses in perfect information games' field (e.g, Alpha Zero in Go), studies have gradually shifted to imperfect information games such as Poker, in which uncertainty management is a key part of this. Reconnaissance blind chess (RBC), a variant of chess, provides a good platform for specialized research on uncertainty management problems, because it has an independent “information sensing” stage. In this paper, we propose Information Entropy of Uncertainty Control (IEUC), which is a new method introduce information entropy theory as the optimization target of uncertainty in imperfect information games. IEUC integrates the whole uncertainty of the environment and strategy evaluation to reduce the uncertainty of the environment more reasonably. We realize a practical RBC agent that incorporates IEUC method. In the experiments, our new agent has shown better effectiveness and robustness than several influential opponents in information sensing, including our previous version that won the champion in terms of uncertainty management in NeurIPS 2019 RBC tournament.