Demonstration-Based Training of Non-Player Character Tactical Behaviors

Abstract

State of the art methods for generating non-player character (NPC) tactical behaviors typically depend on hard-coding actions or minimizing a given objective function. In many games however, it is hard to foresee how the NPC should behave to appear intelligent or to accommodate human player preferences for NPC tactics. In this paper we consider an alternative approach, by training NPC tactical behavior via demonstrations. We propose a heuristic search-based planning method based on previously-developed Experience Graphs, which facilitates the use of behavior demonstration data to plan goal-oriented NPC behavior. Our method provides a principled solution to the problem which tolerates some amount of differences in between the training demonstration and the actual problem and yet still grants guarantees on the quality of the solution output.