Adapting In-Game Agent Behavior by Observation of Players Using Learning Behavior Trees

In this paper we describe Learning Behavior Trees, an extension of the popular game AI scripting technique. Behavior Trees provide an effective way for expert designers to describe complex, in-game agent behaviors. Scripted AI captures human intuition about the structure of behavioral decisions, but suffers from brittleness and lack of the natural variation seen in human players. Learning Behavior Trees are designed by a human designer, but then are trained by observation of players performing the same role, to introduce human-like variation to the decision structure. We show that, using this model, a single hand-designed Behavior Tree can cover a wide variety of player behavior variations in a simplified Massively Multiplayer Online Role-Playing Game