Rares

Rares

Posts by Collection

portfolio

publications

Learning near-optimal decisions: from SAA to robust optimization

Published in In Preparation, 2022

We propose a novel approach that directly recommends robust, near-optimal decisions based on data within the context of contextual stochastic optimization problems.

Recommended citation: Cristian, R., Perakis, G. (2022). Learning Robust Decisions for Contextual Stochastic Optimization Problems Directly from Data

End-to-end learning via constraint-enforcing approximators for linear programs with applications to supply chains

Published in AAAI-23 Main Track, 2022

We present a novel appoach in joint prediction and optimization by introducing a neural network architecture (ProjectNet) capable of approximately solving optimization problems.

Recommended citation: Cristian, R., Harsha, P., Perakis, G., Quanz, B. L., & Spantidakis, I. (2022). End-to-End Learning via Constraint-Enforcing Approximators for Linear Programs with Applications to Supply Chains.

talks

Online semigroup queries on arrays and paths on trees

Published:

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.

Learning near-optimal robust solutions in pricing and beyond

Published:

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.

teaching