The decentralized Wald problem (Q1120536)

The authors discuss a decentralized detection problem: there are two detectors which make independent observations and they decide which of two hypotheses is true. The constant cost c is incurred by each observation and the cost associated with the final decisions \(u_ i\) of the detectors is of form \(J(u_ 1,u_ 2,h)\) where h is the true hypothesis. The detectors' objective is to determine the optimal decision rules \((\gamma_ i,\tau_ i)\) which minimize the expected cost \[ E[c\tau_ 1+c\tau_ 2+J(\gamma_ 1(y_ 1(\tau_ 1)),\quad \gamma_ 2(y_ 2(\tau_ 2)),h)] \] where \(\tau_ i\) is a stopping time, \(y_ i(t)\) is the information obtained by observations of detector i up to the time t, and \(\gamma_ i\) is a decision rule of detector i based on its own information \(y_ i(t)\). It is then shown that the detectors' optimal decision rules are characterized by time-invariant thresholds, and that these thresholds are coupled and their computation requires the solution of two coupled sets of dynamic programming equations. For a typical case an approximate computation of the thresholds is proposed and numerical results are presented.