CAIDAS member Carlo D'Eramo Co-Authors ICML 2025 Spotlight Paper on Advanced Monte-Carlo Tree Search

We are delighted to announce that research co-authored by Carlo D'Eramo, head of the Reinforcement Learning and Computational Decision Making group at CAIDAS, has received significant recognition at the upcoming International Conference on Machine Learning (ICML) 2025. ICML is one of the world's premier and most selective conferences for machine learning research.

The paper, titled "Monte-Carlo Tree Search with Uncertainty Propagation via Optimal Transport", in collaboration among others with Tuan Dam (Hanoi University of Science and Technology) and Odalric-Ambrym Maillard (INRIA Lille), has been accepted with distinction as a spotlight poster, which places the work among the top 2.6% of papers accepted to ICML.

The research introduces an innovative approach to Monte-Carlo Tree Search (MCTS), designed to excel in challenging planning scenarios involving high levels of randomness or incomplete information (stochastic and partially observable Markov decision processes). A key contribution is a novel backup strategy where the uncertainty in value estimates is explicitly modelled using Gaussian distributions and propagated through the search tree using L1-Wasserstein barycenters combined with alpha-divergences. The proposed method comes with strong theoretical guarantees and convincing empirical results showing improvements over established baselines on several challenging benchmark environments.

This prestigious spotlight acceptance highlights the impactful, state-of-the-art research involving Carlo D'Eramo and the quality of work associated with CAIDAS and the University of Würzburg. The selection provides an excellent platform to present these advanced MCTS techniques to the international machine learning community at ICML 2025.

CAIDAS congratulates Carlo and his co-authors on this outstanding achievement!