Monte Carlo Tree Search Planning for Continuous Action and State Spaces (short paper)

Sequential decision-making in real-world environments is an important problem of artificial intelligence and robotics. In the last decade reinforcement learning has provided effective solutions in small and simulated environments but it has also shown some limits on large and real-world domains characterized by continuous state and action spaces. In this work, we aim to evaluate some state-of-the-art algorithms based on Monte Carlo Tree Search planning in continuous state/action spaces and propose a first version of a new algorithm based on action widening. Algorithms are evaluated on a synthetic domain in which the agent aims to control a car through a narrow curve for reaching the goal in the shortest possible time and avoiding the car going off the road. We show that the proposed method outperforms the state-of-the-art techniques.