Research

Github Repos #

DeepCallCenter #

This project is an implementation of the paper:

Dynamic Scheduling of a Multiclass Queue in the Halfin-Whitt Regime: A Computational Approach for High-Dimensional Problems,

authored by Baris Ata and Ebru Kasilarlar.

The paper discusses a dynamic scheduling problem for a call center system. The problem is modeled as a CTMDP, which yields a brownian control problem at the heavy-traffic limit. The paper shows that the brownian control problem can be solved using a deep learning approach, and with that, good control policies can be found for the CTMDP.

Note: This project depends on two of my previous projects:

TorchSimulator #

TorchSimulator is a demonstration framework designed to showcase how parallel, tensorized simulations on GPUs can accelerate performance. It provides a foundation for building custom simulation classes by allowing users to inherit from the provided CTMCSimulator or CTMDPSimulator classes.

In the tutorial.ipynb provided, we walked through a dynamic scheduling problem for a queueing system, where each sample path spans 17 days (approximately 20,000 events).

While simulating 10,000 sample paths takes about 10 minutes on CPU (using C++), GPU acceleration enables running 1,000,000 paths within 15 minutes. This is a 60x speedup.

TorchBSDE Solver Package #

This repository implements Jiequn Han’s DeepBSDE solver in PyTorch. This solver solves high-dimensional PDEs using deep learning.

The code has been restructured to work either as an installable Python package or as a git submodule in other projects. Extensive comments and docstrings have been added to enhance readability and understanding of the implementation.

DeepBSDE Solver Package in TensorFlow (2.0) #

This repository is a fork of Jiequn Han’s DeepBSDE solver implementation, which solves high-dimensional PDEs using deep learning.

The code has been restructured to work either as an installable Python package or as a git submodule in other projects. While changes to the original code are minimal, extensive comments and docstrings have been added to enhance readability and understanding of the implementation.

Publications #

In Preparation #

Ata, B., & Shao, H., Dynamic Capacity Control for Network Revenue Management: A Computational Method Based on Neural Networks
Kleywegt, A. J., Li, Y. & Shao, H., A Markov Decision Process Model for Drivers’ Relocating Behavior in Ride-Hailing Systems

Preprints #

Kleywegt, A. J., & Shao, H. (2022). Revenue Management Under the Markov Chain Choice Model with Joint Price and Assortment Decisions. arXiv:2204.04774
Kleywegt, A. J., & Shao, H. (2021). Optimizing Pricing, Repositioning, En-Route Time, and Idle Time in Ride-Hailing Systems. arXiv:2111.11551
Kleywegt, A. J., & Shao, H. (2020). Tractable Constrained Optimization over Multiple Product Attributes under Discrete Choice Models. arXiv:2007.09193
Kleywegt, A. J., & Shao, H. (2020). Joint Estimation of Discrete Choice Model and Arrival Rate with Unobserved Stock-out Events. arXiv:2003.02313

Journal Papers #

Wang, T., Shao, H., Qu, X., & Eliasson J. (2023). Consolidating Passenger and Freight Transportation in an Urban-Rural Transit System. Fundamental Research
Qu, X., Shao, H., Wang, S., & Wang, Y. (2022). Are More Charging Piles Imperative to Future Electrified Transportation System?. Fundamental Research