Monte-Carlo Tree Search Kaggle Competition on Ludii Games with a $50,000 Prize Pool

We are excited to announce the launch of a Kaggle competition focused on Monte-Carlo Tree Search (MCTS) and games, using the Ludii General Game System.

Kaggle, owned by Google, is an online platform that hosts data science and machine learning competitions, providing a collaborative environment for data scientists, researchers, and enthusiasts to tackle real-world problems. It offers a vast collection of public datasets, tutorials, and a powerful cloud-based environment for data analysis and model development. Kaggle is widely recognised for fostering learning and innovation, with users able to share code, solutions, and insights. It's a valuable resource for both beginners and experts looking to improve their skills and contribute to data science challenges.

The Kaggle competition (available here: https://www.kaggle.com/competitions/um-game-playing-strength-of-mcts-variants):

• A data science competition that tasks you with predicting the win percentage of one MCTS-based agent against another, given string descriptions of both agents and a rich set of features describing the game.
• Runs for approximately three months, starting on September 5th.
• Offers a prize pool of 50,000 USD.

The dataset (available from the competition website):

• Provides outcomes from a variety of MCTS-based agents playing a wide range of board games against one another.
• Includes 1,377 different games/rulesets for public training data, with a distinct set of games reserved as private test data for the competition.
• Features 811 attributes describing each game, including rule descriptions in both English and the Ludii game description language.
• Involves 72 different MCTS agents, incorporating 4 selection strategies, 3 exploration constants, 3 play-out strategies, and an option to toggle Score Bounds.
• Contains 233,000 rows of training data, where each row represents a unique combination. Since each row contains data from multiple plays (due to stochastic outcomes), this dataset represents millions of plays.

Winning entries in the competition will be required to open-source their code. This will allow us to examine high-performing models for new insights and gain a deeper understanding of which MCTS variants excel or struggle across different types of games.

As a reminder, the Ludii General Game Playing (GGP) system (http://ludii.games/) is a highly flexible platform designed for the modelling, play, and analysis of a wide variety of traditional, modern, and newly invented games. Developed as part of the ERC-funded Digital Ludeme Project, and serving as the core software for the GameTable COST Action research network, Ludii employs a unique approach using ludemes—conceptual units that define the rules and components of games. The platform supports various game types, including board games and puzzles, and enables users to create and experiment with new games. With integrated AI agents for automated gameplay and analysis, Ludii is a valuable resource for game researchers, designers, and AI developers.

The organisation of the Kaggle competition is made possible through the collaboration of three universities:

• Dennis J.N.J. Soemers (Maastricht University)
• Éric Piette (UCLouvain)
• Achille Morenville (UCLouvain)
• Matthew Stephenson (Flinders University)
• Kurt Driessens (Maastricht University)
• Mark H.M. Winands (Maastricht University)
• Walter Reade (Kaggle)
• Ashley Chow (Kaggle)