RESEARCH

“Wir müssen wissen. Wir werden wissen."

(We must know. We will know.)

David Hilbert

Here is some of my naive research experience. Many thanks to my professors, who patiently teach a tyro like me :-)

Coming Soon

My professors and I are currently working on dimensionality reduction techniques for optimization problems. Good things may come out soon :-)

Shape Constrained Regression

Supervisor: Prof. Nick Sahinidis @ Georgia Tech

This project aims to find algorithms that are able to do regression on data with a pre-required shape (convex, monotone, etc) and also applications where we can apply shape constraints from prior experience.

I independently studied recently proposed methods based on ADMM, PGD, and ALM, to solve shape-constrained regression problems. To create something novel, I combined ADMM with GPU acceleration (parallel computing) to achieve a speedup of up to 25 times.

I was also introduced to ALAMO developed by Professor Sahinidis, which is a powerful algorithm that adopts both the ideas of symbolic regression and optimization. And I am currently using ALAMO to explore some applications for shape constrained regression and generate physically interpretable models.

Even with the increasing supply of donated organs, the number of patients on the transplant waitlist has also been increasing, creating a significant disparity. Hence, it is important to improve the transplant offer acceptance rates so that fewer organs will be discarded, and more patients can benefit from successful transplants.

In this project, we attempted to build a machine learning model to predict kidney transplant offer acceptances. To make a statistically meaningful and interpretable model, we processed the data on organ transplant offer information from OPTN with Python and did feature engineering, including Gini Importance.

Finally, we proposed a model based on random forest to predict acceptances of organ transplant offers with an accuracy of 86.5% on testing datasets.

Predict Organ Transplant Offer Acceptance

Supervisor: Prof. Pinar Keskinocak @ Georgia Tech

Prostate Cancer Surveillance Strategy Optimization

Supervisor: Prof. Zheng Zhang @ ZJU

Prostate Cancer is common among men. However, it can be deadly without proper treatment. There are usually two ways to detect prostate cancer: Prostate-Specific Antigen test (PSA) and Biopsy test.

PSA is friendly to patients, i.e., the test will not significantly affect a patient's health, or Quality-Adjusted Life Year (QALY), but it is not accurate enough. On the other hand, having a Biopsy test may not be a comfortable experience for a patient; it is painful and can lead to severe health problems. The good news is, Biopsy tests are relatively more accurate. Thus, it is better to combine these two: taking PSA tests regularly and then Biopsy if PSA finds something abnormal.

Our research tried to find a proper threshold probability of getting prostate cancer, which could help decide whether or not to conduct a Biopsy test. Further, we modeled the problem as a 5-state partially observable Markov decision process (POMDP).

After modeling, a possible approach is simulating with transition probabilities and the information matrix showing the probability of getting cancer conditioned on PSA results. Then we can collect QALY values with different conditions and actions. Further, the problem is about optimization.

(But there is another approach to optimize. Reinforcement learning may be a powerful and appropriate tool, and I suppose we will find something interesting in this way.)

Tracking Mask Mandates during COVID-19

Supervisor: Prof. Austin Wright @ UChicago

In Summer 2020, I participated in the Summer Scholar program of the University of Chicago and took part in the data collection work of mask mandates during the COVID-19 pandemic. The database can be found here.

Senior Design: Capstone Project

Supervisor: Prof. Hongwei Wang @ ZJU

There are some environments that are not suitable for humans to perform important tasks, for example, rescuing, exploration in the hazardous chemical leakage scene. Thus, I collaborated with Yue Sun, Shuting Tao, and Xinhao Tong to design an Unmanned Vehicle-Drone Cooperative System.

Our design combined techniques like computer vision, Arduino board programming, and Bluetooth signal processing. And it works in such a way that:

the drone in the air first scans and takes a picture of the whole terrain underneath, then passes the picture to a distant server;
the server processes the picture and comes up with an optimal route planning towards the goal for the unmanned vehicle to follow;
after that, the route is decomposed to several actions including motor running time and rotation degree;
finally, the unmanned vehicle receives the commands wirelessly and moves accordingly

Report Link: https://courses.grainger.illinois.edu/ece445zjui/getfile.asp?id=18986