Modelling Player Behaviour in Total-War Games

在游戏中设计具有挑战性且有趣的敌人是一项艰巨的任务，特别是考虑到现代游戏的复杂性。虽然机器学习是这个问题的一个很有前途的解决方案，但大多数关于学习游戏代理的研究都集中在赢得游戏上。此外，专注于学习乐趣代理的少数尝试通常依赖于乐趣的单一定义，这是有问题的，因为乐趣是什么是主观的。因此，创造乐趣代理的一个潜在解决方案是模拟人类行为，以捕捉乐趣的各种定义。因此，我们建议通过复制人类的内在动机来模拟人类行为。我们认为玩家的动机反映在他的游戏风格中，因此可以被衡量。有了衡量内在动机的指标，我们就可以训练类似人类的代理，优化游戏设计，或者根据玩家的动机实时调整游戏。

Designing enemies in games that are challenging and fun to play against is a difficult task, especially due to the increasing complexity of modern games. While machine learning is a promising solution to this problem, most research on learning game-playing agents focuses on winning the game. Furthermore, the few attempts that focus on learning fun agents often rely on a single definition of fun, which is problematic due to the subjective nature of what fun is.As such, a potential solution for creating fun agents is modelling human behaviour to capture the varied definitions of fun. Therefore, we propose to model human behaviour by replicating their intrinsic motivation. We argue that a player's motivation is reflected in his playstyle and can thus be measured. With metrics for measuring intrinsic motivation, we can train human-like agents, optimize a game's design, or adapt a game to a player's motivation in real-time.