MCTS planners were developed to combat the curse of dimensionality in large state spaces and appear to have a different portfolio of strengths and weaknesses compared to traditional search methods. One example of an algorithm in this class, Upper Confidence applied to Trees (UCT), is credited with providing the foundations for new human-competitive Go-playing programs.

The success in Go motivated further work on understanding the theoretical properties of MCTS algorithms. Several papers were published questioning the tractability of these approaches in the worst case, leading to new algorithms and approaches. While trivial exponential lower bounds were proved, the success of this class of methods in real applications points towards the necessity of developing a better understanding of the type of inputs for which these algorithms perform well. Such an understanding would in turn allow a better recognition of the type of applications for which MCTS algorithms would be best suited.

The goal of this workshop is to understand the techniques that led to the breakthroughs in computer Go, translate these gains into other domains and push forward the theoretical understanding of this probabilistic planning framework. The topics include (but are not limited to):

MCTS planning in MDPs and POMDPs.
MCTS planning in games (Go, RTS games etc.).
Exploration and exploitation in MCTS algorithms.
Bandit-algorithm foundations for MCTS.
Integration of learning and planning.
Representational aspects of MCTS algorithms.
Theoretical foundations of MCTS.
Open problems.
Innovative applications of MCTS algorithms.
Work still in-progress is welcomed to the workshop under the form of position papers (which will be candidates for short, 5 minutes talks) while full-fledged research papers will be considered for long, 20 minutes talks (see submission procedure for details).