Reusing skills for first-time solution of navigation tasks in platform videogames

摘要

We consider the problem of performing real-time navigation in domains where a "god's eye view"is provided. One setting where this challenge arises is in platform videogames, occurring whenever the player wishes to reach an item or powerup on the current screen. Previous agents for these games rely on generating many low-level simulations or training runs for each fixed task. Human players, on the other hand, can solve navigation tasks at a high level by visualising sequences of abstract "skills". Based on this intuition, we introduce a novel planning approach and apply it to Infinite Mario. Despite facing randomly generated, maze-like tasks, our agent is capable of deriving complex plans in real-time, without exploiting precise knowledge of the game's code.