Abstract: Optimizing communication networks such as throughput maximization and power minimization can be viewed as solving a stochastic network optimization problem, which leads to a control algorithm. Generally, the algorithm assumes infinite buffer space and converges to an optimal operating point within O(epsilon^-2) iterations, where epsilon is the proximity to the optimal operating point. In this talk, two aspects of the stochastic network optimization are focused on as the steps toward realization of the technique in practice. In the first part, the control algorithm that can be implemented by using only finite buffer space is considered. Specifically, when each queue in a network has buffer size B, the algorithm achieves within O(e^-B) of the optimality while a delay per queue is O(B) and an average drop rate is bounded by O(e^-B). In the second part, convergence times of a class of algorithms derived from a stochastic network optimization problem with non-convex decision sets are investigated. We show that the algorithm consists of two phases: transient phase and steady-state phase. The transient time, length of the transient phase, is O(epsilon^-1) and O(epsilon^-1.5) under locally-polyhedral and locally-smooth assumptions respectively. Performing a time average of decisions in the steady-state phase leads to faster convergence times that are O(epsilon^-1) and O(epsilon^-1.5) under the aforementioned assumptions.
Bio: Sucha Supittayapornpong is a Ph.D. candidate at University of Southern California, supervised by professor Michael J. Neely. His research interests include stochastic network optimization, distributed algorithms, and convergence analysis, with applications in communication networks, software-defined networking, and machine learning. He completed his M.Eng from Asian Institute of Technology and his B.Eng. from Kasetsart University, Thailand.