Active-Passive Dynamic Consensus Filters With Reduced Information Exchange and Time-Varying Agent Roles

• Published in: IEEE transactions on control systems technology Vol. 28; no. 3; pp. 844 - 856
• Main Authors: Peterson, J. Daniel, Yucelen, Tansel, Sarangapani, Jagannathan, Pasiliao, Eduardo L.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.05.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Aerodynamics; Algorithms; Control algorithms; Control theory; Convergence; Decentralized control; Distributed information systems; Graph theory; Heuristic algorithms; Information exchange; information filtering; Integrals; Multi-agent systems; Multiagent systems; Reagents; Schedules; sensor fusion; sensor network; Symmetric matrices; Vehicle dynamics
• ISSN: 1063-6536; 1558-0865
• DOI: 10.1109/TCST.2019.2896534

Active-passive dynamic consensus filters consist of agents subject to local observations of a process (i.e., active agents) and agents without any observations (i.e., passive agents). The key feature of these filters is that they enable the states of all agents to converge to the average of the observations only sensed by the active agents. Two sweeping generalizations can be made about existing active-passive dynamic consensus filters: 1) they utilize integral action-based distributed control algorithms such that each agent is required to continuously exchange both its current state and integral state information with its neighbors; and 2) they assume that the roles of active and passive agents are fixed ; hence, these roles do not change with respect to time. The contribution of this paper is to introduce and analyze a new class of active-passive dynamic consensus filters using results from graph theory and systems science. Specifically, the proposed filters only require agents to exchange their current state information with neighbors in a simple and isotropic manner to reduce the overall information exchange cost of the network. In addition, we allow the roles of active and passive agents to be time-varying for making these filters suitable for a wide range of multiagent systems applications. We show that the proposed active-passive dynamic consensus filters enable the states of all agents to converge to an user-adjustable neighborhood of the average of the observations sensed by a time-varying set of active agents. We also generalize our results using event-triggered control theory such that agents schedule information exchange dependent on errors exceeding user-defined thresholds ( not continuously). This generalization allows agents to further reduce the overall cost of interagent information exchange and to determine when to broadcast their information to their neighbors thus eliminating the need to synchronize their states. Four illustrative numerical examples and one experimental study are also presented to demonstrate our theoretical findings.