Simulation for Artificial Intelligence Modeling and Assessment

Recent developments in autonomous fighter jets and concepts such as an autonomous wingman are pushing the boundaries for human-autonomy teaming in high-risk military flight operations. Many of these concepts explore the use of aides to accelerate pilot decision-making and reduce cognitive demands. Artificial Intelligence (AI) is a key enabling technology for decision support and automation of flight processes. Machine learning (ML) techniques are the primary method of training and validating modern AI models which requires representative data of increasing size. Acquiring this data is often a major blocker to the development of AI models. This becomes even more challenging when the target domain is aircraft for the U.S. Department of Defense (DoD) where existing datasets may be classified and/or inaccessible. A second requirement in the development of an AI model is an operational environment to integrate, execute, and assess performance in a closed-loop system. The ability to assess the AI safely in a live environment can also be difficult as when technology hasn’t yet fully matured. To address these challenges in the development of AI models for a decision support system, Southwest Research Institute (SwRI) leveraged the U.S. Air Force Research Laboratory’s (AFRL) Advanced Framework for Simulation, Integration, and Modeling (AFSIM) as a solution. This paper explores lessons learned in using AFSIM and its recently added support for the Python programming language to create a testbed for generating data of sufficient size to train Artificial Neural Networks (ANNs) to perform decision support and demonstrated in a closed-loop manner with new/live data.