Application of reinforcement learning to the card game Wizard

This article proposes an application using a reinforcement learning (RL) approach to the card game Wizard. The aim is to create a computer player that is able to learn a winning strategy for the game by himself. Wizard is a partially observable competitive multiplayer game that consists of two game phases, forecasting and trick playing. The biggest challenges in creating a strong player are dealing with multiple rounds which have a different grade of imperfection and the decision on the forecast at the beginning of every game round. We introduce an RL approach to the problem by adopting an existing RL algorithm to the playing phase of the game and by implementing an evaluator of the player's hand card using a Multi-Player-Perceptron to conduct the forecast. The results of our experiments show that the player is able to improve his playing strategy through learning. At the beginning the performance of the learning agent is very bad due to the bad forecasting behavior, but he is able to improve his performance over a few training episodes from 0% won games to approximately 25.68% won games in an experiment with 4 players. Therefore he plays equally strong as his opponents and even outperforms one of them.