Skip to main content Skip to navigation

Reinforcement Learning Environment with LLM-Controlled Adversary in D&D 5th Edition Combat

Project Overview

The document explores the integration of Large Language Models (LLMs) within a Reinforcement Learning (RL) framework, specifically through the lens of Dungeons & Dragons (D&D) 5E combat scenarios, to enhance strategic decision-making in educational contexts. It highlights how LLMs can contribute to AI-driven educational applications by providing deeper insights and enhancing the learning experience in complex environments. The findings reveal that although traditional RL agents have superior performance compared to LLM-controlled adversaries, the strategic depth imparted by LLMs significantly enriches the capabilities of AI in navigating dynamic and unpredictable situations. This suggests that incorporating generative AI into educational tools can foster better engagement and understanding, ultimately improving learning outcomes by leveraging the strengths of both LLMs and RL methodologies.

Key Applications

Reinforcement Learning Environment with LLM-Controlled Adversary in D&D 5E Combat

Context: Educational tool for AI development, particularly in strategic decision-making and adaptive strategies in complex environments

Implementation: Design and implementation of a reinforcement learning environment simulating D&D 5E combat using LLMs to control adversarial agents.

Outcomes: Enhanced strategic decision-making capabilities of RL agents; faster convergence and learning when trained against LLMs.

Challenges: Inconsistent responses from LLMs, slower decision-making processes compared to traditional RL agents.

Implementation Barriers

Technical Barrier

Inconsistent and unstable responses from LLMs can hinder their effectiveness in real-time strategic decision-making.

Proposed Solutions: Implementing prompt strategies to improve stability and response accuracy; integrating LLMs' responses into the RL framework.

Project Team

Joseph Emmanuel DL Dayo

Researcher

Michel Onasis S. Ogbinar

Researcher

Prospero C. Naval Jr

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Joseph Emmanuel DL Dayo, Michel Onasis S. Ogbinar, Prospero C. Naval Jr

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies