Decentralized optimization and multi-agent information fusion

Many emerging applications of embodied machine intelligence require a team of agents to collaboratively construct a shared model by combining their individual observations.  Prominent examples include mapping and localization for multi-agent robotic systems (such as fleets of autonomous ground vehicles or aerial swarms), shared virtual layers for mixed-reality devices (such as AR & VR headsets), and distributed sensing and estimation in sensor networks and internet-of-things devices.  These problems are often formalized as maximum-likelihood or M-estimation (in the case of statistical estimation) or regularized empirical risk minimization (in the case of federated learning), in which case collaborative model estimation reduces to distributed optimization.

While distributed optimization has been an active research area for the last several decades, current state-of-the-art techniques are designed for an operational regime (roughly speaking, scientific computing in data centers) that is radically different from the aforementioned applications.  In particular, standard distributed optimization algorithms often assume that the target problem is unconstrained and convex (neither of which is true for real-world robotics problems), or have iteration or communication complexities that are prohibitively expensive for field robotics systems that must operate in hard real-time while communicating over channels that are frequently low-bandwidth or only sporadically available.

NEURAL is developing novel methods for decentralized optimization and multi-agent information fusion that are specifically adapted to the requirements of embodied machine intelligence applications.  In particular, we are exploring the design of distributed optimization algorithms that are grounded in Riemannian geometry as an elegant approach to coping with nonconvex objectives and constraints.  A primary goal of this work is to devise decentralized methods that preserve the fast convergence of current state-of-the-art centralized techniques (thereby supporting high-precision estimation and real-time execution) while remaining tractable for deployment with limited communication.

Research themes

Optimization, geometry, numerical linear algebra, graph theory

Selected Publications

  • Y. Tian, K. Khosoussi, D.M. Rosen, and J.P. How. “Distributed Certifiably Correct Pose-Graph Optimization”. IEEE Transactions on Robotics 37.6 (Dec. 2021), pp. 2137–2156. DOI.
  • Y. Tian, A. Singh Bedi, A. Koppel, M. Calvo-Fullana, D.M. Rosen, and J.P. How. “Distributed Riemannian Optimization with Lazy Communication for Collaborative Geometric Estimation”. IEEE RSJ / International Conference on Intelligent Robots and Systems. Kyoto, Japan, Oct. 2022. Link.