Robust location parameter estimation in the presence of adversary

Abstract

In classical parameter estimation settings, sensor observation models are often assumed to be known. However, when the sensors themselves become unreliable, the traditional observation models may no longer hold. It is then expected that estimation performance would degrade due to the abnormal behavior of sensor observations. We formulate the estimation problem as a two-person zero-sum game and propose a mini-max estimator with the optimization goal to minimize the worst possible estimation error. We show that there exists a saddle-point solution for a single sensor observation. We then apply our result and characterize the estimation performance for networks with multiple sensors.