A Sample-Efficient Scheme for Channel Resource Allocation in Networked Estimation

Remote estimation over communication channels of limited capacity is an area of research with applications spanning many economically relevant areas, including cyber-physical systems and the Internet of Things. One popular choice of communication/scheduling policies used in remote estimation is the class of event-triggered policies. Typically, an eventtriggering threshold is optimized assuming complete knowledge of the system's underlying probabilistic model. However, this information is seldom available in real-world applications. This paper addresses the learning of an optimal threshold policy based on data samples collected at the sensor. Leveraging symmetry, quasi-convexity, and the method of Kernel density estimation, we propose a data-driven algorithm, which is guaranteed to converge to a globally optimal solution. Moreover, empirical evidence suggests that our algorithm is more sample-efficient than traditional learning approaches based on empirical risk minimization.