FedNPC: A Federated Learning Framework for Large Language Models in Game NPCs
Mengze Hong, Kun Zhang, Shuning Zhang, Zhihang He · 2023
In many modern video games, NPCs often remain unfazed while surrounded by a hail of bullets. The unrealistic behaviors exhibited by game agents can distract from player engagement and have become a significant concern with the recent integration of language models. To address the issue, this paper introduces a conceptual framework (FedNPC) based on centralized Federated Learning, with the objective of enhancing the integrated large language model by enabling NPCs to generate context-appropriate responses through learning from user interactions. This framework takes into account the virtual knowledge of the game and the unique characteristics of individual agents, facilitating model improvement, NPC personalization, and minimizing the impact of undesirable agent removal. The proposed framework is highly practical for implementation and offers valuable insights for industrial practitioners to utilize Federated Learning in game production. Additionally, this research work demonstrates an innovative application of Federated Learning in game development and large language model fine-tuning, highlighting numerous opportunities for further research in this domain.