Prediction of opponent hidden information in imperfect information games

Abstract

Artificial Intelligence (AI) defeated many perfect information games such as checkers, chess, backgammon, othello, scrabble, and so on. In addition the game of Go is the most challengeable perfect information game. Recently, AlphaGo won human champion, Sedol Lee. Through this work, AI can be applied to solve games given perfect information entirely. Next goal of AI is imperfect information game. Major difference between perfect information games and imperfect information games are whether there are hidden information or not. Leading approach of imperfect information games is Nash equilibrium. But, this method has large computational costs in some cases. To overcome this, we propose a new approach which predicts hidden information by using observable data. Information of positions and the kinds of tiles are used as input. If hidden information were perfectly predicted, we don’t need to find Nash equilibrium. Also our model can combine with other existing algorithm using perfect information games. To evaluate our approach, Japanese Mahjong is selected between several imperfect information games. Because, there are usable public data. The prediction accuracy in training is almost 100 percent and the prediction accuracy in test is about 86.5 percent. By using given observable information, three opponent players’ private hand tiles are predicted successfully.