Online and offline learning in multi-objective Monte Carlo Tree Search

Abstract

Multi-Objective optimization has traditionally been applied to manufacturing, engineering or finance, with little impact in games research. However, its application to this field of study may provide interesting results, especially for games that are complex or long enough that long-term planning is not trivial and/or a good level of play depends on balancing several strategies within the game. This paper proposes a new Multi-Objective algorithm based on Monte Carlo Tree Search (MCTS). The algorithm is tested in two different scenarios and its learning capabilities are measured in an online and offline fashion. Additionally, it is compared with a state of the art multi-objective evolutionary algorithm (NSGA-II) and with a previously published Multi-Objective MCTS algorithm. The results show that our proposed algorithm provides similar or better results than other techniques.