Signal Source Localization Using Partially Observable Markov Decision Processes

Abstract

GPS-based systems have the potential to make air travel safer and more efficient, but they are vulnerable to jamming and interference. Localizing signal sources such as jammers can be accomplished cost-effectively using small unmanned aerial vehicles (UAVs). When equipped with a directional antenna, a multirotor UAV can rotate to obtain a bearing measurement to the signal source. This paper presents a control scheme for such a platform, which uses prior observations to determine where new observations should be made. The signal source localization problem is modeled as a partially observable Markov decision process (POMDP). The POMDP formulation allows for a principled approach to optimally solving sequential decision-making problems. Signal source localization problems involve large belief spaces that render such optimal solutions computationally intractable, but recent advances have made approximation methods feasible. This paper presents a POMDP model for signal source localization and explores offline methods to approximate optimal solutions. Simple approximations such as QMDP outperform heuristic strategies if the reward function is carefully selected. More complex solution methods incorporating distribution entropy result in better performance but have greater computational cost.