Hypothesis assignment and partial likelihood averaging for cooperative estimation

Parth Paritosh, Nikolay Atanasov, and Sonia Martínez
Proceedings of the 58th IEEE Int. Conference on Decision and Control, Nice, France, December 2019

Abstract:

We propose a cooperative, decentralized inference algorithm allowing sensor networks to learn a joint parameter best explaining their combined observations. This joint parameter is represented via a probability density over a discrete set of hypotheses. We aim to answer two questions: (i) an agent-hypothesis assignment problem, balancing estimation quality, storage and communication constraints in the networks, and (ii) the design of a provably-correct distributed estimation algorithm on limited hypothesis sets for agents. We make the following contributions to the state of the art. First, our proposed algorithm allows each agent to perform updates on partial likelihoods and exchange local information on a limited hypothesis set, as opposed to the entire hypothesis space. For some of the agents, the limited hypothesis domains may even exclude the true hypothesis. Second, the presented algorithm is the first to not require step-wise renormalization at all hypotheses, while still guaranteeing consensus and convergence of sensor estimates. Third, we also address agent-hypothesis assignment by formulating it as a mixed integer programming problem, that matches agent sub-networks to hypotheses based on a diversity criterion for estimation quality. We provide numerical examples demonstrating the benefits of these algorithms.

File: main.pdf

Bib-tex entry:

@InProceedings{PP-NA-SM:19-cdc,
author = {P. Paritosh and N. Atanasov and S. Mart{\'\i}nez},
title = {Hypothesis assignment and partial likelihood averaging for cooperative estimation<},
booktitle = {58th IEEE International Conference on Decision and Control},
pages = {},
year = {2019},
address = {Nice, France},
month = {December}
}