Abstract

The objective of this study is to design and implement a reinforcementlearning (RL) environment using D\&D 5E combat scenarios to challenge smallerRL agents through interaction with a robust adversarial agent controlled byadvanced Large Language Models (LLMs) like GPT-4o and LLaMA 3 8B. This researchemploys Deep Q-Networks (DQN) for the smaller agents, creating a testbed forstrategic AI development that also serves as an educational tool by simulatingdynamic and unpredictable combat scenarios. We successfully integratedsophisticated language models into the RL framework, enhancing strategicdecision-making processes. Our results indicate that while RL agents generallyoutperform LLM-controlled adversaries in standard metrics, the strategic depthprovided by LLMs significantly enhances the overall AI capabilities in thiscomplex, rule-based setting. The novelty of our approach and its implicationsfor mastering intricate environments and developing adaptive strategies arediscussed, alongside potential innovations in AI-driven interactivesimulations. This paper aims to demonstrate how integrating LLMs can createmore robust and adaptable AI systems, providing valuable insights for furtherresearch and educational applications.