Chebyshev metric based multi-objective Monte Carlo tree search for combat simulations

Monte Carlo Tree Search (MCTS) is frequently used for online planning and decision making in large space problems, where the move maximizing a reward score is chosen as the optimal solution. As many problems have more than one objective, this paper presents a multi-objective version of MCTS. The algorithm employs a non-linear scalarization function, the Chebyshev metric based function, as a basis for node expansion and final move selection. We first test it in a multi-objective benchmark problem, the Deep Sea Treasure, to show its performance. Then how it is applied for commander behavior modeling in an infantry combat simulation is described. Results from a series of comparative experiments suggest that the performance of the presented approach matches the state of the art multi-objective MCTS implementation, and more suitable for the infantry commander decision making than the linear MCTS.