Intelligent, Fast Reinforcement Learning for ISR Tasking (IFRIT)

Scientific Systems Company, Inc. (SSCI) in conjunction with our academic partners at MIT, propose the Intelligent, Fast Reinforcement Learning for ISR Tasking (IFRIT) system, to provide sophisticated autonomous end to end ISR mission execution capability. The IFRIT framework will enable real time adaptive mission planning that responds effectively to dynamic environments, leveraging reinforcement learning (RL) and Macro-Action Partially Observable Markov Decision Process (Mac-POMDP) decision techniques. Leveraging these RL techniques in conjunction with macro-actions integrating SSCI’s CMA behavior software for EO/IR area search and RF emitter geolocation, provides flight tested genetic algorithms and other combinatorial techniques for online, real-time multi-objective optimization, fusing live sensors, operator inputs, and comms quality data. A mission planner utilizing RL, Mac-POMDP decision techniques and CMA behaviors provides a mission planner that performs effective planning of macro-actions for an ISR Mission. This mission planner augmented with modern embedded CPU/GPU combinations, including GPUs for deep network usage computation to assist with rapid RL policy approximations makes IFRIT an effective mechanism for mission planning and dynamic replanning as external factors present in a real time dynamic environment. Incorporating a mission monitor for output of this mission planner provides the IFRIT system with models to ensure desirable behavior and tune the level of autonomous replanning desired by the operator.