Decentralized sequential decision making with asynchronous communication

Decentralized sequential decision making with asynchronous communication Georgios Fellouris We consider three statistical problems – hypothesis testing, change detection and parameter estimation– when the relevant information is acquired sequentially by remote sensors; all sensors transmit quantized versions of their observations to a central processor, which is called fusion center and is responsible for making the final decision. Under this decentralized setup, the challenge is to choose a quantization rule at the sensors and a fusion center policy that will rely only on the transmitted quantized messages. We suggest that the sensors transmit messages at stopping times of their observed filtrations, inducing in that way asynchronous communication between sensors and fusion center. Based on such communication schemes, we propose fusion center policies that mimic the corresponding optimal centralized policies. We prove that the resulting decentralized schemes are asymptotically optimal under different statistical models for the observations. These asymptotic optimality properties require moderate, or even rare, communication between sensors and fusion center, which is a very desirable characteristic in applications.