Real-Time Path Planning using a Simulation-Based Markov Decision Process

Abstract

This paper introduces a novel path planning technique called MCRT which is aimed at non-deterministic, partially known, real-time domains populated with dynamically moving obstacles, such as might be found in a real-time strategy (RTS) game. The technique combines an efficient form of Monte-Carlo tree search with the randomized exploration capabilities of rapidly exploring random tree (RRT) planning. The main innovation of MCRT is in incrementally building an RRT structure with a collision-sensitive reward function, and then re-using it to efficiently solve multiple, sequential goals. We have implemented the technique in MCRT-planner, a program which solves non-deterministic path planning problems in imperfect information RTS games, and evaluated it in comparison to four other state of the art techniques. Planners embedding each technique were applied to a typical RTS game and evaluated using the game score and the planning cost. The empirical evidence demonstrates the success of MCRT-planner.