LLM-based wargame scenario generation with domain ontology and ECA rules

Abstract

Rapid advancements in the text generation capabilities of large language models (LLMs) have expanded their applicability across various domains, notably in scenario generation. This paper introduces a dynamic framework that utilizes the advanced text generation of LLMs to enhance scenario generation and decision-making in the wargame scenario. Our framework, fine-tuned on domain-specific data, automates the generation of complex and adaptive scenarios, responding to user interactions. It integrates a wargame domain ontology to ensure scenario accuracy and employs event–condition–action (ECA) rules for decision-making to improve the realism and explainability of scenarios. In addition, through direct preference optimization (DPO), the framework continually refines scenarios based on user feedback, thereby ensuring highly sophisticated simulations. This approach not only diminishes the time and resources needed for wargame scenario generation but also significantly boosts the effectiveness of training. Our findings highlight the transformative potential of LLMs in automating scenario generation and decision support for wargame simulations, thereby broadening the applicability of scenario generation research to various fields.