Wargaming plays a crucial role in defence preparedness, but often the huge volume of outputs it generates can be impenetrable, even for the most experienced data processing team.
We've demonstrated how large language models (LLMs) can solve that challenge by turning complex wargaming output data into easy to use, secure information that improves the scenario interrogation and analysis.
In a programme funded by the Defence, Science and Technology Laboratory (Dstl), our detailed research has shown there is great potential to use LLMs in wargaming and significantly reduce the burden on the operator. It also complements our work on Red Force Response (friendly forces), where AI-trained adversaries were used to demonstrate possible responses to a Blue Force (enemy forces) Course of Action in wargaming simulations, further cementing our expertise in this area of military intelligence.
The promise of LLMs in wargaming
LLMs are known for their ability to summarise complex data through their text processing and generating capabilities. They can analyse and asses large data sets from a variety of sources far quicker than any manual approach. This is ideal for Command: Modern Operations (CMO), an advanced, high-fidelity wargaming simulation platform that produces large volumes of complex data on completion of a given scenario.
Our research scrutinised whether an LLM could be used reliably and securely to interrogate the output of a CMO scenario, for example, a complex multi-domain engagement involving sea, air and land units. Could it help the analyst more easily understand the result of a battlefield scenario and the key factors that drove it?
The answer is a resounding yes. Multiple technologies were considered, including combining Retrieval Augmented Generation (RAG) with a local LLM. RAG is a state-of-the-art technique that allows use-case specific data in everyday formats such as PDF, CSV or XML so it can be easily included in the context for an LLM response. Though less powerful than online counterparts like Chat GPT, local LLMs provide greater privacy and data control.
A set of possible use cases were provided and tested across two phases of the six- month project. We also created a robust framework tool for quantifying the accuracy and reliability of the LLM generated information.
Overwhelmingly our studies showed that, used in the right way, LLMs can helpfully interrogate and disseminate output information of complex wargaming scenarios. This strengthens the training benefits, vastly reduces operator burden, and improves resilience and preparedness.
Even better, the techniques can be developed flexibly around changing components such as data types, tools, methodologies and evaluation metrics, ensuring this new approach can evolve with ever-changing demands and challenges.
Read more on this in our published paper here>> Using AI in Wargaming Simulation as a Multi-Domain Decision Support Tool
