Conference Presentation, 2022 INFORMS Annual Meeting, Indianapolis, IN
We develop a method for producing robust decisions in predict-optimize tasks. In particular, we introduce two notions of robustness: (1) decisions which minimize maximum cost with respect to noise/uncertainty in the objective, (2) producing stable decisions which do not change significantly under perturbation to the training data.
Conference Presentation, 2022 INFORMS Annual Meeting, Indianapolis, IN
We developed a novel neural network architecture that can learn to approximately solve optimization problems. We incorporate this network into an end-to-end framework which can effectively learn forecasts producing near-optimal decisions.
Seminar, MIT Operations Research Center Student Seminar, Cambridge, MA
We consider a common class of database range queries which consists of evaluating a given function on contiguous subranges of arrays. For instance, this may be the average or minimum of a range. Such queries are also common subroutines in various algorithms. A common approach to tackling the range query problem for semigroup operators is to precompute the answer for a small subset of ranges, and combine these solutions when answering any given query. For instance, the sum of the elements from index i to index j can be computed as the answers to the ranges from i to k and k+1 to j for i <= k < j. This introduces an inherent tradeoff between the precomputation complexity and the query complexity - the more precomputation, the less query time required. Moreover, we consider a data-driven case and design a method to make better precompuation decisions than traditional methods.