Post-doc researcher, UC Berkeley
Distributed optimization with limited computation and communication resources
In a large-scale network, reliable communication and powerful local computation are always two key factors for a successful implementation of any distributed optimization algorithms. However in practice, large-scale networks may suffer from different limitations on communication and local computation power. In this work, we study inexact splitting methods and utilize them to solve distributed optimization problems in presence of communication and computation errors. In the first part, we consider the communication constraint with a limited communication data-rate, and propose two distributed optimization algorithms with an iteratively refining quantization based on inexact proximal gradient method. It is shown that if the parameters of the quantizers satisfy certain conditions, then the quantization error decreases linearly and the convergence of the distributed algorithms is guaranteed. In the second part of this work, we focus on developing new robust and stochastic distributed optimization algorithms with inexact local updates. We propose a new inexact splitting method called inexact fast alternating minimization algorithm to solve distributed optimization problems, allowing for local computation and communication errors. We derive complexity upper-bounds on the number of iterations for the proposed algorithm and provide sufficient conditions on the errors for the convergence of the algorithm. In the end, we demonstrate the proposed distributed optimization algorithms by solving a distributed optimal control problem originated from a real building control application.
Ye Pu is currently a postdoctoral researcher (Swiss National Science Foundation – Early Postdoc. Mobility Fellowship) in the Hybrid Systems Lab at the University of California, Berkeley.
She received a B.S. degree and a M.S degree for Electrical Engineering from Shanghai Jiao Tong University, China, in 2008 and the Technical University Berlin, Germany, in 2011, and a Ph.D degree from Swiss Federal Institute of Technology in Lausanne, Switzerland, for her work “Splitting methods for distributed optimization and control” in 2016.
Her research focuses on optimization-based control (model predictive control), optimization algorithms and machine learning with the application of power networks.