Seminar - Dr. Sergio Rodriguez, Applied Scientist, Amazon
Wednesday, March 11, 2020 - 3:30am to 4:30am
Title: Batch Bayesian Optimization with Noisy Constraints
Constrained Bayesian optimization (BO) is aimed to solve constrained optimization problems where objective and constraint(s) are black-box expensive-to-evaluate functions. In this talk, I develop a constrained batch BO algorithm, dubbed c-BBO, which allows the user to: (i) incorporate noise in objective and constraints, (ii) suggest new querying points in batch, and (iii) specify the tolerance to constraint violation at each time step. First, I introduce the main components of the c-BBO algorithm, namely, the underlying Bayesian model for objective and constraint(s) based on Gaussian Processes regression, and our data acquisition mechanism based on a novel extension of the constrained Expected Improvement. Then, I illustrate c-BBO using several synthetic examples showing that c-BBO manages to sample at feasible locations within the user pre-specified tolerance, and that it is able to find the feasible optimum in a considerably small number of steps (compared to approaches that use a single update). Finally, I briefly outline how c-BBO can be utilized to tune ranking systems. This work was jointly developed with Dr. Aditya Maheshwari.
Dr. Sergio Rodriguez obtained his PhD in Statistics and Applied Probability from UCSB under the supervision of Prof. Mike Ludkovski. He also holds two masters degrees: Statistics, and Electrical and Computer Engineering (with emphasis in Control and Signal Processing) both from UCSB. Currently, he is an Applied Scientist at Amazon.com, where he is responsible for developing on-line learning methods in the context of search advertising. His research interests lie in the intersection of machine learning, on-line learning/sequential design, and optimization. In his current role, he has been focusing on Bayesian optimization (BO) and its extensions: noisy constrained BO, batch querying, and multi-task BO problems (i.e. BO with the aid of a simulator). He is also very passionate about implementation of the latter techniques using free programming languages, so he has recently been working closely with the main contributors of BO routines in python (such as Emukit).