1 / 14

Accelerating Approximate Consensus with Self-Organizing Overlays

This paper proposes a method called PLD-Consensus that accelerates the convergence of Laplacian-based consensus algorithms by creating self-organizing overlays. The approach biases the consensus process to trade precision for speed, making it suitable for spatial computing applications such as flocking, swarming, and collective decision making. The PLD-Consensus algorithm is shown to converge in O(diam) time, with robustness to mobility and parameter variations.

ellisj
Download Presentation

Accelerating Approximate Consensus with Self-Organizing Overlays

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Accelerating Approximate Consensus with Self-Organizing Overlays Jacob Beal 2013 Spatial Computing Workshop @ AAMAS May, 2013

  2. PLD-Consensus: With asymmetry from a self-organizing overlay… … we can cheaply trade precision for speed.

  3. Motivation: approximate consensus Many spatial computing applications: • Flocking & swarming • Localization • Collective decision • etc … Current Laplacian-based approach: • Sensor fusion • Modular robotics • DMIMO Many nice properties: exponential convergence, highly robust, etc. But there is a catch…

  4. Problem: Laplacian consensus too slow On spatial computers: • Converge O(diam2) • Stability constraint  very bad constant Root issue: symmetry [Elhage & Beal ‘10]

  5. Approach: shrink effective diameter • Self-organize an overlay network, linking all devices to a random regional “leaders” • Regions grow, until one covers whole network • Every device blends values with “leader” as well as neighbors In effect, randomly biasing consensus Cost: variation in converged value

  6. Self-organizing overlay Creating one overlay neighbor per device: • Random “dominance” from 1/f noise • Decrease dominance over space and time • Values flow down dominance gradient

  7. Power-Law-Driven Consensus Asymmetric overlay update Laplacian update

  8. Race: PLD-Consensus vs. Laplacian

  9. Analytic Sketch • O(1) time for 1/f noise to generate a dominating value at some device D [probably] • O(diam) for a dominance to cover network • Then O(1) to converge to D±δ, for any given α (blending converges exponentially) Thus: O(diam) convergence

  10. Simulation Results: Convergence 1000 devices randomly on square, 15 expected neighbors; α=0.01, β=0, ε=0.01] Much faster than Laplacian consensus, even without spatial correlations! O(diameter), but with a high constant.

  11. Simulation Results: Mobility No impact even from relatively high mobility (possible acceleration when very fast)

  12. Simulation Results: Blend Rates β irrelevant: Laplacian term simply cannot move information fast enough to affect result

  13. Current Weaknesses • Variability of consensus value: • OK for collective choice, not for error-rejection • Possible mitigations: • Feedback “up” dominance gradient • Hierarchical consensus • Fundamental robustness / speed / variance tradeoff? • Leader Locking: • OK for 1-shot consensus, not for long-term • Possible mitigations: • Assuming network dimension (elongated distributions?) • Upper bound from diameter estimation

  14. Contributions • PLD-Consensus converges to approximate consensus in O(diam) rather than O(diam2) • Convergence is robust to mobility, parameters • Future directions: • Formal proofs • Mitigating variability, leader locking • Putting fast consensus to use…

More Related