Objective and Summary

On June 16th I defended my PhD in machine learning with a research focus on reinforcement learning and visualization. I am currently looking for post-graduate employment following work at NASA Ames and the XPRIZE Foundation.

My grad school career is summarized below. The works come in four distinct areas. First, I developed a simulator for fire, vegetation growth, timber sales, and weather. Second, I developed visual analytic tools for exploring the policy space of Markov Decision Processes (MDPs), including the wildfire simulator. Since many MDPs are defined by computationally expensive simulators, I next developed a surrogate modeling method that brings interactive specification of policy, reward, and optimization functions to large state space Markov Decision Processes. My final area of focus is Bayesian policy search using the surrogate model I developed.

Outside my doctoral research, I am also very active in the development of open source code and organizations. I developed a federally-recognized 501c3, the Privly Foundation, dedicated to online privacy education. The foundation's activities include developing Open Source software, technology workshops, and supervising student developers.

For a full professional history, please view my LinkedIn profile.

Papers

Visualizing High-Dimensional MDPs with Model-Free Monte Carlo

McGregor, S., Houtman, R., Montgomery, C., Metoyer, R., & Dietterich, T. G.

Best Paper Nomination

This work extends Model-Free Monte Carlo to large state space Markov Decision Processes (MDPs) by factoring the state space. We successfully generate surrogate trajectories for a large state space MDP (more than 10^1000000 states) with 8 state features. We posted an extended version of the paper to arXiv.

Paper Citation BibTeX

Fast Optimization of Wildfire Suppression Policies with SMAC

McGregor, S., Houtman, R., Montgomery, C., Metoyer, R., & Dietterich, T. G.

This paper is the first application of random forests for direct policy search. Additionally, it is the first case of optimizing wildfire suppression policies across 100 year time spans.

Paper Citation

Factoring Exogenous State for Model-Free Monte Carlo

McGregor, S., Houtman, R., Montgomery, C., Metoyer, R., & Dietterich, T. G.

This work shows how to reduce the simulation time of a high-dimensional ecological Markov Decision Process (MDP) from 7 hours to less than a seconds. We apply the method outlined in the RLDM paper.

Fast Simulation for Computational Sustainability Sequential Decision Making Problems

McGregor, S., Houtman, R., Buckingham, H., Montgomery, C., Metoyer, R., & Dietterich, T. G.,

This paper, presentation, and poster give preliminary results for modeling the wildfire policy MDP with Model-Free Monte Carlo.

Paper Poster Presentation Slides Citation BibTeX

Allowing a Wildfire to Burn: Estimating the Effect on Future Fire Suppression Costs

Houtman, R. M., Montgomery, C. A., Gagnon, A. R., Calkin, D. E., Dietterich, T. G., McGregor, S., & Crowley, M.

This paper outlines the wildfire simulator I collaboratively developed for evaluating wildfire suppression decisions. Wildfires can reduce timber harvests in the present, but suppression decisions must balance firefighting costs, reduced fuel risks post-fire, and timber revenues.

Paper Citation BibTeX

Interactive visualization for testing Markov Decision Processes: MDPVIS

McGregor, S., Buckingham, H., Dietterich, T. G., Houtman, R., Montgomery, C., & Metoyer, R.

The development of a computationally expensive MDP simulator demonstrated the absence of adequate tooling for sequential decision making problems. This paper presents a visual analytic system, MDPvis, that is easy to interface with MDP simulators and optimizers and extensible with additional visualizations.

Demo Paper Citation BibTeX

Toward Visualization Methods for Interactive Improvement of MDP Specifications

McGregor, S., Dietterich, T. G., & Metoyer, R.

This workshop paper presents an early version of MDPvis and discusses the need for stronger visualizations in the reinforcement learning community.

Paper Presentation Slides Citation BibTeX

MDPvis: An Interactive Visualization for Testing Markov Decision Processes

McGregor, S., Buckingham, H., Houtman, R., Montgomery, C., Metoyer, R., & Dietterich, T. G.

This paper is an extended abstract of the paper presented at the IEEE Symposium on Visual Languages and Human-Centric Computing.

Paper Presentation Slides Citation BibTeX

Facilitating Testing and Debugging of Markov Decision Processes with Interactive Visualization

McGregor, S., Buckingham, H., Dietterich, T. G., Houtman, R., Montgomery, C., & Metoyer, R.

This paper introduces Markov Decision Processes to visualization researchers as a theoretical formulation for sequential optimization problems. We explain the MDP formulation, present a visual analytic system exploring simulator-defined MDPs, and evaluate the system on a high dimensional MDP domain.

Paper Presentation Slides Citation BibTeX

Presentations

Beneficial AI

United Nations, International Telecommunication Union

I served as panelist and/or rapporteur for this gathering of UN Agencies, NGOs, and academics. The group effort centered on developing guidlines and principles for the beneficial use of Artificial Intelligence. As part of my panel participation I drafted three documents in collaboration with panel moderators. These include one document on beneficial AI with the head of the Montreal Institute for Learning Algorithms (MILA), another document on privacy and security with the director of the United Nations Interregional Crime and Justice Research Institute, and a final document on measuring the positive and negative impacts of AI technologies with the head of UNICEF's venture fund.

Citation

Solving Grand Challenges with Artificial Intelligence

McGregor, S.

I keynoted the Thrival Innovation Festival on the topic of Artificial Intelligence for good. This talk was to a general business and startup audience and centered around the teams competing in the IBM Watson AI XPRIZE. Following the keynote I sat on a panel with David Danks (CMU Philosophy department chair), Amrit Dhir (Google for Entrepreneurs), Jana Eggers (Nara Logics), and Iba Masood (TARA.AI).

Citation

Interdisciplinary Collaboration: Finding new Ways of Extracting Scientific Insights from Solar Observations.

McGregor, S., Munoz-Jaramillo, A., Dhuri, D., & Berea, A.

I spent the summer of 2017 working collaboratively with computer scientists and heliophysicists to predict solar flares from raw Solar Dynamics Observatory images. This presentation was an invited talk to present our preliminary results to the NASA community.

My contributions to this research included defining the software architecture for heliophysics deep learning research, including visualization methods for extracting scientific insights from trained neural network models.

Since the development process was unusual by NASA standards, the first quarter of the presentation focused on our style of agile development methods as a means of conducting interdisciplinary research.

Citation BibTeX

How to Encrypt Your Content on Any Website

McGregor, S., Karve, S., & Davidson, J.

This presentation explained the technical foundations of the priv.ly project.

Citation BibTeX

Making Your Privacy Software Usable

Davidson, J., & McGregor, S.

I collaborated with UX researcher Jen Davidson on a presentation of usability evaluation methods and privacy software ethics.

Citation BibTeX

The Open Privacy Stack: Privly

McGregor, S., & Davidson, J.

OSCON is the premiere Open Source industry conference. This presentation focuses on our work on the priv.ly project's implementation.

Presentation Slides Citation BibTeX

Privacy (or not) in a Digital Age

McGregor, S.

This talk covered threats to electronic privacy and mitigation strategies. Topics covered included encryption, browser extensions, networking, and malware.

Citation BibTeX

Posters

Incorporating Future Values into Analysis of Current Wildfires

McGregor, S., Houtman, R., Metoyer, R., Dietterich, T. G., Montgomery, C., & Crowley M.

Optimization and forestry are not often combined. We created this poster to gather input from forestry practitioners regarding the reports and visualizations that are useful in their work. The idea is to present our computer science research in a way that will lead to its adoption in a field that emphasizes field work over simulation students.

Poster Citation BibTeX

Service

Coalition of Graduate Employees

The Coalition of Graduate Employees (CGE) is a labor union representing more than 1,600 graduate employees. During my time in graduate school I helped double the size of the bargaining unit and budget, bargain a multi-million dollar contract, collaborate with university human resources on the implementation of our health plan, serve as ranking member of our delegation to the state federation, and serve as secretary-treasurer.

EECS Graduate Student Association

Despite having more than 400 graduate students in the department, I am the only one in my 6+ years to have organized events for the whole grad student population. I recently formed a student association and recruited a slate of officers to continue building on the social events I organized.

Reviewer

I reviewed reinforcement learning papers for this top-tier conference.

Privly Foundation

The Privly Foundation is a side project I kickstarted during my second year of graduate school. It is unrelated to my graduate studies, but has seen significant successes, including:

  • Designed and built method for posting private content to third-party social media: The Priv.ly Project
  • Gained attention in The Atlantic, Der Spiegel, Hacker News, and numerous other outlets.
  • Obtained 501(c)(3) tax exemption
  • Organization has mentored ten students in the Google Summer of Code (2013-2015)
  • Initiated online privacy meetup group in Portland, Oregon featuring presentations from leaders in security, incident response, citizen journalism, censorship circumvention, and cryptography
  • Funded at 260 percent of a Kickstarter campaign and raised total financial and in-kind support in excess of $100,000