Decentralized Decision-Making for Multi-Agent Networks: the State-Dependent Case
Joni Shaska, Urbashi Mitra · 2021 IEEE Global Communications Conference (GLOBECOM) · 2021
We consider a new formulation of the decentralized detection problem with parallel agent configuration. In particular, each agent in the network exists in a set of pre-specified states that affects the distribution of their observations as well as the underlying hypothesis. As such, observations are conditionally dependent. Following a person-by-person design methodology, it is shown that the Bayes optimal detection rule for each agent is a likelihood ratio test with a state dependent threshold. Moreover, it is shown that even for statistically identical agents, the optimal rules for the agents may not be the same. Motivated by this, we turn our attention to large networks and find the error exponent, and show that as the number of agents increases there is no loss of asymptotic optimality if the agents use the same rule, dramatically reducing the complexity of computing the decision rules for each agent.