[1I4-OS-31a] OS-31

[1I4-OS-31a-04] Building Symbolic Game AI using LLM

Action evaluations with large language models and learning of evaluation functions

〇Motoharu Kano¹, Naoki Hamada¹, Takayuki Shimotomai¹ (1. KLab inc)

Keywords:Large Language Models, GameAI, Symbolic AI

The AI (character AI) used to control non-player characters (NPCs) in entertainment games is required to behave as determined by the developer, rather than pursuing the best actions according to the game rules. Therefore, the current mainstream of character AI is to use symbolic AI methods such as behavior trees to make it easier for developers to understand and control behavior. However, when introducing reinforcement learning to game developments, there may be issues such as the difficulty of alignment and changes in the game environment during the development. In this study, to solve this problem, we propose a new behavior control method that combines symbolic AI and LLM. First, the NPC's specifications, in-game situation, and behavior are expressed in text, and behavior evaluation data is generated by evaluating these using LLM. The generated evaluation data is used as annotation data to build small-scale machine learning models that can run in real time. In the experiment, we prepared behavior rules and policies for the character AI in an actual game, created a model using this method, and validated that the AI can select actions based on the requirements set in the game.