Planning for Non-Player Characters by Learning From Demonstration

John Drake, University of Pennsylvania


In video games, state of the art non-player character (NPC) behavior generation typically depends on hard-coding NPC actions. In many game situations however, it is hard to foresee how an NPC should behave to appear intelligent or to accommodate human preferences for NPC behavior. We advocate the creation of a more flexible method to allow players (and developers) to train NPCs to execute novel behaviors which are not hard-coded. In particular, we investigate search-based planning approaches using demonstration to guide the search through high-dimensional spaces that represent the full state of the game. To this end, we developed the Training Graph heuristic, an extension of the Experience Graph heuristic, that guides a search smoothly and effectively even when a demonstration is unreachable in the search space, and ensures that more of the demonstrations are utilized to better train the NPC's behavior. To deal with variance in the initial conditions of such planning problems, we have developed heuristics in the Multi-Heuristic A* framework to adapt demonstration trace data to new problems. We evaluate our approach in the Creation Engine game engine by modifying The Elder Scrolls V: Skyrim (Skyrim) to accommodate our NPC behavior generators and experiments. In Skyrim, players are given "quests" which are composed of several objectives. NPCs in the game sometimes accompany the player on quests, but state-of-the-art companion NPC AI is not sophisticated enough to behave according to arbitrary player desires. We hope that our work will lead to the creation of trainable NPC AI. This will enable novel gameplay mechanics for video game players and may augment video game production by allowing developers to train NPCs instead of hard-coding complex behaviors.

Subject Area

Computer science

Recommended Citation

Drake, John, "Planning for Non-Player Characters by Learning From Demonstration" (2018). Dissertations available from ProQuest. AAI10749752.