On Multisensor Activation Policies for Bernoulli Tracking

In this work, we propose a family of sensor-activation policies coupled with a learning method for information-seeking sensor activation in multisensor Bernoulli tracking applications. Sensor activation, and sensor management more generally, are of great interest in multi-agent networks and Internet of Things (IoT) applications, where limited energy, sensing, and communication resources have to be efficiently allocated for optimal inference. Non-myopic control, resource constraints, and the partial observability of the Bernoulli target are the main challenges addressed in this work. The novelty of our approach is threefold. First, a belief-space Markov decision process (MDP) reformulation is proposed for the Bernoulli tracking and control problem that incorporates uncertainties in both object existence and object state. Second, a parametric family of sensor-activation distributions is proposed as control policies that harness the mutual information between the sensor measurements and the belief state. Thirdly, a novel reward metric is employed to capture the information gain on the Bernoulli belief state of a specific combination of sensor activations. A Bayesian actor-critic (BAC) reinforcement learning (RL) methodology is employed to further refine the policy by maximizing a discounted reward over an infinite horizon and under an imposed activation constraint. Numerical simulations validate our approach and show an improved tracking performance over a uniform sensor-activation method.