Reinforcement learning-based dynamic scheduling for threat evaluation

A novel reinforcement learning-based sensor scan optimisation scheme is presented for the purpose of multi-target tracking and threat evaluation from helicopter platforms. Reinforcement learning is an unsupervised learning technique that has been shown to be effective in highly dynamic and noisy environments. The problem is made suitable for the use of reinforcement learning by its casting into a “sensor scheduling” framework. An innovative action exploration policy utilising a Gibbs distribution is shown to improve agent performance over a more conventional random action selection policy. The efficiency of the proposed architecture in terms of the prioritisation of targets is illustrated via simulation examples.