Online Learning Techniques for Space Situational Awareness (Poster)

The Sensor Allocation Problem (SAP) is a stochastic control problem that can rarely be solved to optimality. Real world systems often use heuristic solutions in place of optimal solutions when optimal solutions are not computationally feasible. This work is a step toward developing an approach to solve the SAP in the Space Situational Awareness (SSA) domain using Deep Reinforcement Learning (DRL) instead of a traditional heuristic approach. DRL Agents learn and adapt from experience and thus may demonstrate favorable characteristics in dynamic environments over traditional heuristic solutions. We develop a methodology to apply DRL algorithms to problem classes for which good heuristics exist. As a proof of concept of our general approach, we apply it to an under-actuated non-linear control problem. The problem is solved with a traditional heuristic approach. We use that solution to train a DRL agent to reach a similar level of performance. Performance characteristics of each solution, as well as to that of traditional DRL training approaches, are compared. The long-term goal is to extend our approach to more complex problem classes, and ultimately to provide improved solutions to the SAP in the SSA domain.