METRON INCORPORATED — Department of Defense STTR Phase I: AF21B-T001

Description

We will develop a theoretical foundation for all-domain Course of Action (COA) planning and scheduling leveraging AI/ML and automation.  This Edge Based Cross-domain Services (EBCS) framework will simplify COA planning by drawing inferences about margins of value-to-cost.  Predicting how satisfaction of high-level objectives and constraints scales with the number of entities, distance, and service time-scales.  This framework requires context to draw inference, we will adapt the game StarCraft II’s functionality to apply reinforcement learning to train sophisticated agents strategy, then apply this RL to more complex problems like degrading an IADS.   We will use a framework of theories, primarily, Deep Learning with Graph Neural Networks. We will leverage advances by Dr. Gombolay’s deep learning team with graph attention networks for learning hierarchical coordination strategies for multi-agent teams.  This prior work leverages heterogeneous graph neural networks, which process graphical models of multi-agent scheduling problems through graph convolutional and attention operators to learn high-quality, computationally-efficient policies for scheduling agents under temporal and resources constraints.   EBSC COA generation benefits include hours saved/simplified COA planning.  This solution can be adapted to optimize a major airline’s flight segment planning and scheduling by integrating it with numerous alliance airline partner plans and schedules to find efficiencies.