Abstract
This work examines a cross-layered caching problem for distributed estimation in wireless sensor networks (WSNs). In WSNs, large amounts of data are produced continuously over time, and storing all the data collected from the sensors can be costly. In distributed estimation applications, sensors first gather information about a common phenomenon, and then forward the information to a fusion center where the final estimate is computed. By assuming that the parameters are correlated over time, the estimation quality at the fusion center can be improved by combining both present and past information, where the latter can be obtained from cached data. Different from conventional caching problems, where the goal is to reconstruct the sensors' observations, our caching strategy is designed to minimize the long term average mean-square error (MSE) of the final estimate. This problem can be modelled as a Markov decision process but, due to the curse of dimensionality, is solved here using a greedy one-step-ahead caching strategy, which only minimizes the expected MSE in the next time slot. This results in a nonlinear fractional programming problem that is solved approximately using semi-definite relaxation and a modified Dinkelbach's algorithm. The effectiveness of the proposed scheme is demonstrated through numerical simulations.