Automated Planning for Pathfinding in Real-Time Strategy Games
Naveed, Munir Hussain
This thesis is focused on the design of a new path planning algorithm to solve path planning problems in dynamic, partially observable and real-time environments such as Real-Time Strategy (RTS) games. The emphasis is put on fast action selection motivating the use of Monte-Carlo planning techniques. Three main contributions are presented in this thesis. The first contribution is a Monte-Carlo planning technique, called MCRT, that performs selective action sampling and limits how many times a particular state-action pair is explored to balance the trade-off between exploration of new actions and exploitation of the current best action. The thesis also presents two variations of MCT as the second contribution. The first variation of MCRT randomly selects an action as a sample at each state seen during the look-ahead search. The second variation, called MCRT-CAS, performs the selective action sampling using corridors. The third contribution is the design of four real-time path planners that exploit MCRT and its variations to solve path planning problems in real-time. Three of these planners are empirically evaluated using four standard pathfinding benchmarks (and over 1000 instances). Performance of these three planners is compared against two recent rival algorithms (Real-time D*-Lite (RTD) and Local Search Space-Learning Real-Time A* (LSS-LRTA)). These rival algorithms are based on real-time heuristic search. The results show that a variation of MOCART, called MOCART-CAS, performs action selection significantly faster than the rival planners. The fourth planner, called the MG-MOCART planner, is evaluated using a typical Real-Time Strategy game. The MG-MOCART planner can solve the path planning problems with multiple goals. This planner is compared against four rivals: Upper Con�dence bounds applied to Trees (UCT), LSS-LRTA, Real-Time Dynamic Programming (RTDP) and a rapidly-exploring random tree (RRT) planner. The performance is measured using score and planning cost. The results show that the MG-MOCART planner performs better than its rival techniques with respect to score and planning cost.