Date/Time
Date(s) - February 05, 2019
11:45 am - 12:35 pm
Location
CSE 122
Categories
Fatemeh Hashemi, Ph.D.
Instructor & Postdoctoral Scholar
Grado Department of Industrial & Systems Engineering, Virginia Tech
Abstract: Sampling-Controlled Search Methods and Applications
Simulation Optimization (SO) has recently generated much attention in various fields. Such interest is primarily because of SO’s flexibility, allowing the implicit specification of functions within the optimization problem, thereby providing the ability to embed virtually any level of complexity. However, most of the existing methodologies for solving SO problems require arbitrarily large data acquisition, and hence not practical in rapid decision making contexts with expensive/streaming data.
The focus of this talk centers on the design of sampling analytics for solving SO problems that leads to the most economical use of expensive data both in theory and practice. Our proposed Adaptive Sampling Controlled Stochastic Recursions (A-SCSRs) are distinct in that they operate as online parameter estimation routines for near-instantaneous classification of streaming data. A real example is the problem of decoding ElectroEncephaloGraph (EEG) data from human brain of a paralyzed patient, towards constructing neurosprosthetics. The key idea in devising A-SCSRs is a novel sequential sampling scheme to trade-off sampling effort against statistical error, so as to learn and sample from data parsimoniously as it evolves through the parameter space. ASCSR’s convergence rate is provably optimal in a well-defined rigorous sense. Amply evident in numerical experiments, the efficiency of our parsimonious sampling scheme results in dramatically better efficiency rates among prominent competing machine learning methods. We also talk about A-SCSR’s potential adoption in fields as varied as disease diagnosis, renewable energy systems, and operations management systems, where rapid decision-making is required owing to the urgent nature of the questions posed.