Title: Bayesian Congestion Control over a Markovian Network Bandwidth Process: A multiperiod Newsvendor Problem
I present a Bayesian congestion control problem in which a data source must select the transmission rate over a network whose available bandwidth is modeled as a time homogeneous ﬁnite-state Markov Chain. The decision to transmit at a rate below the instantaneous available bandwidth results in an under-utilization of the resource while transmission at rates higher than the available bandwidth results in a linear penalty. The trade-off is further complicated by the asymmetry in the information acquisition process: transmission rates that happen to be larger than the instantaneous available bandwidth result in perfect observation of the state of the bandwidth process. In contrast, when transmission rate is below the instantaneous available bandwidth, only a (potentially rather loose) lower bound on the available bandwidth is revealed. We show that the problem of maximizing the throughput of the source while avoiding congestion loss can be expressed as a Partially Observable Markov Decision Process (POMDP). We prove structural results providing bounds on the optimal actions. The obtained bounds yield tractable sub-optimal solutions that are shown via simulations to perform well.
The general form of this problem is mathematically identical to the Newsvendor problem in Operations Research, where the demand is a Markovan process and the decision-maker must select the number of items to store (inventory level) in order to maximize the total expected reward.
This is joint work with Prof. Bhaskar Krishnamachari and Prof.Tara Javidi from UCSD.
Bio: Parisa Mansourifard received her Bacholar of Science and Master of Science in Electrical Engineering from Sharif University of Technology in 2008 and 2010, respectively. She joined University of Southern California in 2011 with Viterbi fellowship where she is currently pursuing a Ph.D. degree. Her research interests include learning theory, intersection of optimization and learning, stochastic resource allocation, and intersection of operations/management science and engineering.