Neuro-evolutionary approach to multi-objective optimization in one-player mahjong

Abstract

In Mahjong, there are several objectives at play to win the game: to (1) win early, (2) gain a large number of points, (3) avoid losing points, and (4) prevent other players from winning. These objectives often conflict with each other, thereby creating tradeoffs. In this research, we make a evaluating function of Mahjong as a multiobjective optimization. Further, Mahjong requires multimodal behavior, particularly because the first-place player must win early and avoid losing points by the last-place player gaining a large number of points in the final hand. In this paper, we propose an evaluation function for Mahjong in the form of a multi-objective optimization problem. Modular multi-objective neuro-evolution of augmenting topologies (MM-NEAT) is a framework for evolving modular neural networks in which each module defines a separate policy. Evolution optimize these policies and when to use them. Given the above, we focus on two objectives in one-player Mahjong: to (1) win early and (2) gain a large number of points. We also verify the effectiveness of MM-NEAT for Mahjong.