Can the Best Jass AIs Beat the Top Humans?

The performance of Artificial Intelligence (AI) in imperfect information games is not at its peak. In the case of the Swiss card game Jass, previous work showed that the best bot at the time could compete with active amateur human players with over 10 years of experience on average[6]. Since the human vs. AI experiments in the previous paper are scarce, these will be the focus of this paper to continue the research on AI in cooperation games. The agent used implements a Determinized Monte Carlo Tree Search (DMCTS) algorithm for the card selection and a Deep Neural Network (DNN) for the trump selection. In this paper, we first look at the current state of research according cooperation games and the previous state according the Jass game. Our main contributions are the implementation of an optimised Graphical User Interface (GUI) and the connection with the currently strongest Jass agent. Furthermore, we provide details according to the conducted experiments such as setup, performance and limitations to help future researchers. The sobering truth of the research question is that the best Jass agent cannot consistently win against well-practiced Jass teams. Potential reasons for this and a statement about the agent’s performance level are described in this paper.