Richard Linares, University of Minnesota, Roberto Furfaro, University of Arizona
Keywords: Sensor Tasking, Deep Reinforcement Learning
Abstract:
The field of Space Situational Awareness (SSA) has progressed over the last few decades with new sensors coming online, the development of new approaches for making observations, and new algorithms for processing them. Although there has been success in the development of new approaches, a missing piece is the translation of SSA goals to sensors and resource allocation; otherwise known as the Sensor Management Problem (SMP). This work solves the SMP using an artificial intelligence approach called Deep Reinforcement Learning (DRL). Stable methods for training DRL approaches based on neural networks exist, but most of these approaches are not suitable for high dimensional systems. The Asynchronous Advantage Actor-Critic (A3C) method is a recently developed and effective approach for high dimensional systems, and this work leverages these results and applies this approach to decision making in SSA. The decision space for the SSA problems can be high dimensional, even for tasking of a single telescope. Since the number of SOs in space is relatively high, each sensor will have a large number of possible actions at a given time. Therefore, efficient DRL approaches are required when solving the SMP for SSA. This work develops a A3C based method for DRL applied to SSA sensor tasking. One of the key benefits of DRL approaches is the ability to handle high dimensional data. For example DRL methods have been applied to image processing for the autonomous car application. For example, a 256×256 RGB image has 196608 parameters (256*256*3=196608) which is very high dimensional, and deep learning approaches routinely take images like this as inputs. Therefore, when applied to the whole catalog the DRL approach offers the ability to solve this high dimensional problem. This work has the potential to, for the first time, solve the non-myopic sensor tasking problem for the whole SO catalog (over 22,000 objects) providing a truly revolutionary result.
Date of Conference: September 19-22, 2017
Track: Tasking