Learning to Play Jass: Human-like AI with Supervised and Deep Learning

This project explores how artificial intelligence can learn to play the Swiss card game Jass in a human-like manner. Using supervised learning and Wasserstein GANs, we aim to model expert-level strategies and evaluate imitation quality through multiple metrics.

Initial Situation

The game of Jass presents a unique challenge in AI development due to its team-based dynamics, partial observability, and strict rule set. While rule-based bots exist, they often lack flexibility and do not capture the nuances of expert human play. Modern deep learning techniques, particularly from reinforcement and generative learning, provide new opportunities for more realistic and adaptable AI players.

Problem Statement / Project Goal

The main goal of this project is to develop an AI system that not only plays Jass effectively but also mimics the behavior of human players. We focus on supervised learning from gameplay logs on the website schieber.ch, followed by adversarial imitation learning using Wasserstein GANs. Evaluation includes accuracy metrics, UMAP-based distribution comparison, and gameplay testing in the Zühlke Jass Bot Challenge environment.

Solution developed and its benefits

To address the challenge of creating a human-like Jass-playing AI, we developed a hybrid pipeline that combines supervised learning and adversarial training. The model learns from expert gameplay logs and refines its behavior using a Wasserstein GAN (WGAN) to better imitate the distribution of human decisions. The solution was integrated into the Zühlke Jass Bot Challenge environment to validate its practical performance.

• Accurate imitation of expert players through supervised learning on thousands of real Jass moves.
• Human-like behavior generation via adversarial training with a WGAN architecture.
• Improved training stability and diversity of actions thanks to Wasserstein loss and gradient penalty.
• Legal move handling without needing post-hoc masking, integrated into the game environment.
• UMAP visualizations to compare model behavior with human distributions.
• Evaluation via win rate comparisons and NLL-based imitation metrics.
• Modular architecture for plugging into the official Jass Bot Challenge framework.
• Reusable codebase for further research in imperfect information games.

Compared to previous rule-based bots, this solution achieves a more nuanced playing style, showing an understanding of team dynamics and subtle in-game patterns. By training on human data and refining strategies through adversarial imitation, the AI avoids overly mechanical moves and approximates the richness of real gameplay.

The system also serves as a strong baseline for future research in team-based card games, enabling further experiments with belief modeling, reinforcement learning, or transformer-based architectures.

Key terms

• JassA traditional Swiss trick-taking card game, known for its strategic complexity and variants like Schieber and Coiffeur.
• AIArtificial Intelligence, used in this context to develop bots that can play Jass by learning strategies and making decisions.
• Supervised LearningA machine learning approach where models are trained on labeled data (e.g., past Jass games) to imitate human decision-making.
• GANGenerative Adversarial Network, a framework with two competing neural networks (generator and discriminator) used to model complex distributions like human play behavior.
• Wasserstein GANA more stable variant of GAN that uses Wasserstein distance to measure divergence between generated and real data distributions, improving training stability and realism.
• Imitation LearningLearning policies by mimicking expert human gameplay, typically used in your supervised phase.
• GAILGenerative Adversarial Imitation Learning, an adversarial approach to imitation learning, useful for making the AI more human-like.
• Jass Bot ChallengeThe official environment provided by Zühlke, used to run and evaluate your models.

Customer

Optobyte AG
Werner Bättig
Bölli 188
5600 Lenzburg
https://www.optobyte.ch/

Team