SF-MCTS: Score Feedback Monte Carlo Tree Search for Digital Curling in Continuous State Space

Deep Reinforcement Learning has been utilized in digital curling, a sport that requires experience and tactics. Nevertheless, current curling agents still difficult to accurately handle curling games, because the continuous state and action space of curling lead to the loss of position information during discretization. In this paper, we have designed a new curling agent based on curling rules. A Curling Location Extraction Policy-Value Network (LE-PV Network) has been introduced to address the issue of information loss in state space. Additionally, a Score Feedback Monte Carlo Tree Search (SF-MCTS) has been proposed to enhance the value function. It can solve the problem of discretization of action space and improve the decision-making capabilities of digital curling algorithms. Our method has demonstrated superior performance compared to existing models in experiments.