Neuro-Symbolic Creation of Non-Playable Characters: An ablation study of the synergy of a Knowledge Graph in a Reinforcement Learning model.

Abstract

This thesis explores the integration of Knowledge Graphs (KG) with Reinforcement Learning (RL) to enhance the adaptability and efficiency of Non-Playable Characters (NPCs) in dynamic video game environments. By combining KG and RL in a neuro-symbolic approach, the study aims to address the lengthy and resource-intensive training processes typically required for RL agents. Using an ablation methodology, the research tests the effectiveness of various KG complexities, including partial and fully dynamic KGs, in a custom-built simulation environment employing Python, PyTorch, and Pygame. While results are pending, the study expects to demonstrate that KG-RL integration can significantly reduce training time and improve NPC adaptability. This research may offer broader implications for developing AI agents in various dynamic systems, laying foundational insights into the scalability and applicability of the KG-RL model.