Dynamic Sensor Scan Optimisation Using Reinforcement Learning

This paper presents the application of reinforcement learning-an unsupervised learning technique-to sensor scan optimisation for the purpose of multi-target tracking from helicopter platforms. The sensor considered in this paper is the forward looking IR (FLIR) sensor. The potential for reinforcement learning based optimisation techniques to provide improved performance over deterministic scan pattern search methods is illustrated by way of simulation examples. Ground targets to be tracked by FLIR are modelled as Markovian processes. The optimisation problem is cast into a "sensor scheduling" framework to facilitate the use of reinforcement learning. Prioritisation of targets in multiple-target detection problems and targets of different threat levels are also considered and illustrated with simulation examples.