Constrained control of non-playing characters using Monte Carlo Tree Search

In this paper, we apply the Monte Carlo Tree Search (MCTS) method for controlling at once several virtual characters in a 3D multi-player learning game. The MCTS is used as a search algorithm to explore a search space where every potential solution reflects a specific state of the game environment. Functions representing the interaction abilities of each character are provided to the algorithm to leap from one state to another. We show that the MCTS algorithm successfully manages to plan the actions for several virtual characters in a synchronized fashion, from the initial state to one or more desirable end states. Besides, we demonstrate the ability of this algorithm to fulfill a specific requirement of a learning game AI : guiding the non player characters to follow a predefined and constrained learning scenario and, if necessary, to adapt their decision to unexpected events in the simulation.