Curling Tactics Analysis with Reinforcement Learning


Loading...

Author / Producer

Date

2025

Publication Type

Master Thesis

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

Curling is often referred to as "Chess on Ice" due to its tactical complexity. Featuring a continuous state and action space along with a complex, stochastic state transition, curling poses significant challenges for analytical modeling. In this thesis, we develop a self-supervised approach to learning curling tactics by combining backward induction with an actor-critic reinforcement learning method. Leveraging the finite horizon of the game, we apply backward induction to reduce the computational complexity. Each stage is solved using the deep deterministic policy gradient (DDPG) method to learn both a policy (actor) and a Q-function (critic). The best learned policies demonstrate consistent performance in a simulated tournament, especially when playing with the advantage of the last rock. Beyond policy learning, the Q-function approximation enables a quantitative evaluation of different tactical options, e.g., in an after-game review or during training.

Publication status

published

External links

Editor

Contributors

Examiner : Cederle, Matteo

Book title

Journal / series

Volume

Pages / Article No.

Publisher

ETH Zurich

Event

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

09478 - Dörfler, Florian / Dörfler, Florian check_circle

Notes

Funding

Related publications and datasets