Research

Adaptive Rank Sampling with Robust Solution for Assortment Planning

Published in Manuscript, under revision, 2019

Research work with Professor Patrick Jaillet (MIT Operations Research Center) and Dr. Mai Anh (Singapore-MIT Alliance). In this paper, we show connections between parametric and rank-based choice models. We propose a new approach to sample ranks and update their distribution from population. We then propose a data-driven robust optimization model, i.e., likelihood robust optimization, for non-parametric assortment planning and we show how to solve the robust model in a tractable way. We provide experimental results using a real-like retail dataset, which shows the efficiency of our rank sampling approach and the tractability of our robust method.

Download here

Consecutive Optimizer for XGBoost

Published in draft, 2019

Research work with Professor Teo Chung-Piaw in NUS Business School. We found that the optimization goal of the XGBoost algorithm well matches the conditions of consecutive optimizer. Based on this connection, We purpose a refined XGBoost algorithm that utilize the info from consecutive tree leaves to reduce the over-fitting issue. This algorithm has a O(KT[mn^2+dlogn]) time complexity that is tractable for prediction task of size 10-100K.

Download here

Lower Confidence Bound Policy for Optimization bounded Regression

Published in preparation, 2019

We study the problem where prediction objectives are some parameters in an optimization problem. We purpose a lower confidence bound policy such that the loss of optimization by predicting the parameters wrong has a tight bound. Moreoever, we purpose a tuning techiqnue such that the prediction by out method substantially outperforms than normal regression. We apply this algorithm to knapsack and network follow optmization problem, deriving an average of 2% decrement in expected optimization loss.