Asymptotic Performance In Heterogeneous Human-machine Inference Networks

We analyze the asymptotic performance of likelihood ratio based collaborative human-machine decision making systems. Human agents are assumed to make threshold based local binary decisions, where the thresholds are considered as random variables. The proposed hybrid system consists of multiple human sub-populations, with the thresholds of each sub-population characterized by non-identically distributed random variables, and a limited number of machines (physical sensors) whose exact values of thresholds are known. For such a hybrid system, we derive the asymptotic performance at the fusion center in terms of Chernoff information. When available, the effect of side information for human sensors is also studied in this paper. Moreover, a budget-constrained human worker selection optimization problem is formulated to determine the optimal number of workers selected from each sub-population according to different costs so that the best decision making performance could be achieved.