Main Subcommittee Category
Emerging Concepts and Innovative TechnologiesHow would you label your submission?
IndustryAbstract
The Department of Defense (DoD) is committed to enhancing the interconnectivity of Artificial Intelligence (AI) and platforms for Joint All Domain Command and Control (JADC2), requesting $1.8 billion for these capabilities in 2024. This is forecast to increase substantially as the budget requests for 2025 and 2026 are being developed, since the DoD is looking for more opportunities for increasing integration between platforms. Traditionally, the DoD’s focus on specific platforms has led to development of capable platforms that are incapable of communicating without significant redesign and engineering efforts. Many have tried to solve this interoperability gap by developing global standards that platforms should implement, however, as standards have been developed, they have evolved, leading to the same lack of interoperability to exist between disparate systems. Others, such as the Defense Advanced Research Project Agency (DARPA), have developed technologies enabling rapid integration between systems, one of which is known as the System of Systems Toolchain for the Integration of Heterogeneous Electronic Systems (STITCHES). While this technology is useful for rapid integration, the learning curve for its Domain Specific Language (DSL) can be difficult for those who have no functional programming experience. How can DoD organizations increase their integration velocity while leveraging the latest tools available from DARPA and taking advantage of the latest developments in the field of AI?
One solution is to employ Large Language Models (LLMs) to facilitate the development of STITCHES DSL translation code. In 2022, OpenAI opened many possibilities with ChatGPT. Many open source LLMs have now been developed, some specifically trained to write correct code. This paper will explore the viability of utilizing an open source LLM to effectively construct STITCHES transforms between publicly available standards by developing a dataset based on the DSL, training/fine-tuning it, and finally, explain how our LLM performed on various prompts.