An efficient Monte Carlo approach for optimizing communication constrained decentralized estimation networks

We consider the design problem of a decentralized estimation network under communication constraints. The underlying low capacity links are modeled by introducing a directed acyclic graph where each node corresponds to a sensor platform. The operation of the platforms are constrained by the graph such that each node, based on its measurement and incoming messages from parents, produces a local estimate and outgoing messages to children. A Bayesian risk that captures both estimation error penalty and cost of communications, e.g. due to consumption of the limited resource of energy, together with constraining the feasible set of strategies by the graph, yields a rigorous problem definition. We adopt an iterative solution that converges to an optimal strategy in a person-by-person sense previously proposed for decentralized detection networks under a team theoretic investigation. Provided that some reasonable assumptions hold, the solution admits a message passing interpretation exhibiting linear complexity in the number of nodes. However, the corresponding expressions in the estimation setting contain integral operators with no closed form solutions in general. We propose particle representations and approximate computational schemes through Monte Carlo methods in order not to compromise model accuracy and achieve an optimization method which results in an approximation to an optimal strategy for decentralized estimation networks under communication constraints. Through an example, we present a quantification of the trade-off between the estimation accuracy and the cost of communications where the former degrades as the later is increased.