Parley as an approach to distributed detection

摘要

Decentralized detection networks implement local likelihood ratio tests at each sensor and combine the local decisions into a global decision. In general this is a “one-pass” procedure, with data transmission from the sensors (or subordinate decision makers) to the fusion center. There have been treatments of multipass schemes, but usually the approach has been to update the local decisions based on the previous fused decision. We explore the use of feedback (of all sensor decisions to all sensors) and successive retesting and rebroadcasting of the updated decisions until a common decision or consensus is reached, an operation we call parley due to its similarity to the process of discussion and negotiation by a team of human decision makers. We consider two modes of operation of such a network. In the first, all sensors are as correct as possible at all times. We see that this network is fast in reaching consensus, but not particularly as “correct” as it could be. Under the second scheme, we observe that via feedback it is possible for the network to make an optimum decision-this may, of course, require several rounds of parley. Of particular interest is that under both schemes a consensus always occurs-the sensors never “agree to disagree”