A Reinforcement Learning Approach for Multisensor Tracking

We introduce a framework for learning a policy to track a target in a multisensor scenario. By using deep reinforcement learning in conjunction with POMDPs, the scenario itself does not have to be made known to the agent that controls the sensor network. This is a distinct difference to other state of the art algorithms adapting to the environment. To illustrate the described methods and algorithms, an example consisting of two sensors illuminating a common area with an opaque obstacle is presented. Finally, the proposed RL approach is compared to a state of the art cognitive approach, where we can show that it performance similarly or up to 50% better.