Dan Regan, Ball Aerospace
Keywords: Deep Learning, Neural Network, Modular, SSA, Tasking, Simulation, Autonomous, Optimized, Scalable, Resilience, Adapable
Abstract:
Modular neural networks are evaluated for their ability to automatedly task an SSA system. For this study, an optical, steerable SSA system is modeled within a MATLAB simulation and its behavior is automated by a modular neural network. This network-of-networks approach begins with the development of three artificial neural network blocks designed to perform specific tasks essential to the execution of the SSA mission – Search, the discovery of new objects; Reacquisition, the subsequent measurement of an object to refine its state model; and Tasking, the decision to Search or Reacquire at a specific time given the state of the system. These network blocks are trained individually through deep Q reinforcement learning to optimize reward values, which have been designed to maximize the performance of the respective functions. The Tasking, Search, and Reacquisition blocks are then integrated together to form a modular neural network architecture capable of executing a basic SSA mission, with individually optimized functional components. The integrated network can then be retrained as a whole through transfer learning to demonstrate this architecture’s ability to adapt and re-optimize to changing mission performance metrics and system configurations. The advantages of this modality are explored in its adaptability and cost savings over traditional, hard-coded algorithms approaches by obviating a man-in-the-loop.
Date of Conference: September 11-14, 2018
Track: Space Situational Awareness