Short-Term Scientific Mission

Efficient MCTS Variants for Stochastic Environments and Simultaneous Action Spaces

Main theme: Optimisation Under Uncertainty
Grantee: Jakub Kowalski, Uniwersytet Wrocławski, Wrocław, Poland
Host: Dennis J. N. J. Soemers, Maastricht University, Maastricht, Netherlands

Start date: 2026-03-23
End date: 2026-06-19
Awarded: 2026-02-19
Report approved: 2026-07-28

Description

The goal of this STSM is to advance and evaluate Monte Carlo Tree Search (MCTS) methods for solving sequential decision problems in highly uncertain adversarial environments. The plan is to formalize the potential usefulness of low-sample determinization of world states in highly nondeterministic domains and to refine algorithmic variants for simultaneous-move games under strict computational constraints. Research activities will involve the implementation and analysis of specialized algorithm enhancements. These methods will be tested across multiple benchmark domains, including complex tabletop game environments that will require the development of programming-friendly interfaces to facilitate standardized evaluation.

Achievements

AI search algorithms used in game-playing and decision-making systems involving uncertainty and imperfect information were improved through collaboration between the grantee and local experts. Working with Prof. Mark Winands and Dr. Dennis Soemers, new methods were developed to help AI systems make smarter decisions under uncertainty; in particular, dynamic resource allocation enhancements were proposed for Ensemble Determinization MCTS and validated on a number of popular tabletop games. Progress was also made on two further research projects, and a bachelor’s thesis and an internship were supervised. Beyond the concrete outputs, lasting connections were built between the two institutions, with further joint projects and funding applications now planned.