Sensor Scheduling for Target Tracking Using Approximate Dynamic Programming

To trade off tracking accuracy and interception risk in a multi-sensor multi-target tracking context, we study the sensor-scheduling problem where we aim to assign sensors to observe targets over time. Our problem is formulated as a partially observable Markov decision process, and this formulation is applied to develop a non-myopic sensor-scheduling scheme. We resort to extended Kalman filtering for information-state estimation and use unscented transformation for trajectory sampling in order to reduce the number of samples required for Q-value approximation. We make decision using a simulation-based approximate dynamic programming method called policy rollout, which is implemented by means of receding horizon control. The effectiveness of our approach is substantiated through an example in which multiple sensors are deployed to track a single target. Key-Words: Non-myopic sensor scheduling, Partially observable Markov decision process, Interception risk, Policy rollout, Unscented transformation