Sensor tasking utilizing deep reinforcement learning in a random finite set framework

Access full-text files




Shah, Akhil Kanu

Journal Title

Journal ISSN

Volume Title



There is a growing need to increase the capabilities of existing sensor arrays to monitor a large amount of space objects orbiting the Earth with a limited number of opportunities to observe these objects. Due to geopolitical considerations and financial cost, it is infeasible to create an array of sensors that can monitor each space object and accurately describe its state. Instead of brute force techniques by increasing the number of sensors worldwide, the current advancements in computational capability along with new algorithms for multi-target filtering and reinforcement learning has allowed a pathway to begin solving the non-myopic, heterogenous sensor tasking problem. This work employs the labeled multi-Bernoulli filter in conjunction with advanced, deep reinforcement learning techniques such as the policy gradient Q-learning algorithm and deep Q-networks. The filter and reinforcement learning techniqures are used together to track ten targets in geosynchronous orbit, while a linear Kalman filter and the reinforcement learning techniques are used to evaluate their effectiveness in multi-agent learning scenarios. The future deployment of these algorithms and their specific logistical considerations are also discussed with potential solutions.


LCSH Subject Headings