Graduate Students Seminar

Session Chair:   Madison Christ
Discussant:        Dr. Rathinam

Speaker 1:  Nathan Tamiru

Title: High-Dimensional Bayesian Optimization for Expensive Black Boxes: A case in Traffic Calibration 

Abstract
Many problems in science and engineering require optimizing a function that can only be evaluated through an expensive simulation: there are no gradients, the output is noisy, and the budget is a few hundred evaluations. Bayesian optimization (BO) handles this by fitting a Gaussian process surrogate to the evaluations made so far and using its uncertainty to decide where to evaluate next. BO is well understood in low dimensions, but whether its advantage survives as the dimension grows is much less clear. This talk builds GP-based BO from first principles and then tests it on a recent benchmark that calibrates origin–destination travel demand for real San Jose freeway networks against measured traffic counts using the SUMO traffic simulator.

Speaker 2: Pallabi Ghosh 

Title: Causal Effect Estimation for Covariates with Measurement Error

Abstract
Covariate measurement error presents an important challenge to causal effect estimation in observational studies due to the violation of the unconfoundedness assumption. Ignoring measurement error and using naïve propensity scores lead to biased estimates of the average treatment effect (ATE). This presentation provides a brief overview of the propensity-score framework and two approaches that account for covariate measurement error. The first, developed by Battistin and Chesher, uses a small-measurement-error approximation to characterize and correct treatment-effect bias. The second, proposed by McCaffrey, Lockwood, and Setodji, constructs weighting functions whose conditional expectations equal the inverse propensity scores based on the true covariates. The presentation also discusses our current research using a joint finite-mixture model with latent propensity-score subclasses to estimate treatment effects.