📝 SummarXiv

Abstract

We consider a real-time tracking system where a binary Markov source is monitored by two heterogeneous sensors. Upon command, sensors send their observations to a remote sink over error-prone channels. We assume each sensor exhibits state-dependent detection accuracy and may occasionally fail to detect the source state. At most one sensor is scheduled for sampling at each time slot. We assess the effectiveness of data communication using a generic distortion function that captures the end application's objective. We derive optimal sink-side command policies to minimize the weighted sum of distortion and transmission costs. To model the uncertainty introduced by sensing failures (of the sensors) and packet loss, we formulate the problem as a partially observable Markov decision process (POMDP), which we then cast into a belief-MDP. Since the belief evolves continuously, the belief space is discretized into a finite grid and the belief value is quantized to the nearest grid point after each update. This formulation leads to a finite-state MDP problem, which is solved using the relative value iteration algorithm (RVIA). Simulation results demonstrate that the proposed policy significantly outperforms benchmark strategies and highlights the importance of accounting for state-dependent sensing reliability in sensor scheduling.