Scott Sussex
I'm a fourth year undergraduate in Applied Mathematics at Harvard. I'm a member of the Statistical Reinforcement Learning Lab,
led by Professor Susan Murphy. I've worked on a project to create a model that produces individualized forecasts in mobile health studies, and more recently
a project on batch off-policy reinforcement learning. This summer I will be at the Robotics Institute of Carnegie Mellon University doing
research in John Dolan's lab, focusing on autonomous vehicles.
Projects I’ve Worked On
Stitched Trajectories for Off-Policy Learning
This work aims to reduce the variance of batch off-policy estimators by using the Markov property to construct new trajectories from those in the original dataset. We demonstrated empirically that our method could be applied to a range of relevent importance sampling methods to reduce the variance of the estimators they produce, and proved it mathematically for ordinary importance sampling. It began as a class project for Professor Susan Murphy’s graduate seminar on sequential decision making. The project poster I presented at the end of the class is available here. I’ve continued working on this project since. It’s a collaboration between myself, Omer Gottesman and Professor Finale Doshi-Velez, along with Yao Liu and Professor Emma Brunskill at Stanford.
Information Elicitation for Credit Services
I worked with Harvard graduate student Eric Mibuari on the problem of using community reports to forecast a borrower’s probability of repaying a loan. We designed an incentive compatible mechanism to obtain reports from a borrower’s social circle on the borrower’s creditworthiness, and formulated a lender’s optimization problem, where the lender seeks to maximize profits whilst varying the number of community members they seek reports from and the amount of cash incentive these people are paid for reporting truthfully. This project was completed as part of Professor Yiling Chen’s graduate seminar on topics at the interface of computer science and economics.
Musical Accompaniment to Solo Translation via RNN Encoder-Decoder
This work was done as part of Professor Alexander Rush’s graduate machine learning class. I worked with two other undergraduates on applying research in machine translation to the problem of music composition. We tested our model on the task of jazz music composition. In a human trial, subjects could not distinguish between MIDI files produced by our model compared to MIDI files of a transcription of pieces by famous jazz soloists.