Christopher Salomon Smith

About Me

I am a first-year Computer Science PhD student at Stony Brook University co-advised by Professors Erez Zadok and Omkant Pandey. My research interests include distributed systems, security, and reinforcement learning (specifically multi-armed bandits). Other interests include proof assistants like Lean and Coq, and various math topics including algebra, measure theory, and most recently topology.

I am originally an Arizona native born and raised in the city of Chandler. Prior to joining the PhD program, I also completed my undergraduate degree in Computer Science and Mathematics at Stony Brook.

Feel free to contact me: csasmith (at) cs (dot) stonybrook (dot) edu

Projects

Decentralized Multi-Armed Bandit Can Outperform Classic Upper Confidence Bound

The multi-armed bandit problem is a special case of online reinforcement learning where an agent faced with a set of actions tries to balance the task of learning the action with the greatest expected reward against the task of maximizing cumulative reward from actions played so far. The distributed case often has several agents cooperating to solve the same bandit problem, with applications in recommendation systems, medical trials, robotics, and more. This work presents an algorithmic design for cooperative bandit problems that adapts two state-of-the-art single-agent "Upper Confidence Bound" (UCB) algorithms, proves it outperforms the single-agent benchmark, and most importantly is completely decentralized. To date, every other cooperative bandit algorithm uses some global knowledge of the network to achieve its performance, which is not always realistic given the size and nature of agent networks that may appear in practice.

I was primarily responsible for implementation, conducting comprehensive simulations, and communicating results in the paper.

Congressional Redistricting Tool

Designed and implemented a full stack web application for a capstone project in a four-person team that provides the user with a selection of congressional districting plans generated on a HPC cluster by a Markov chain Monte Carlo method. Users can visually compare districting statistics and geometry, and then optimize a chosen districting for equal population variance between districts via a novel simulated annealing algorithm run on the application server in real time that works with highly granular census data. Such a tool could be used by legislators to efficiently search for more equitable districting layouts.

I was primarily responsible for the application design, React front end, and Java Spring + Hibernate backend interfacing with a MySQL database. Our tool produced significant improvements across all states in population equality variance during the simulated annealing stage that is currently unattainable by existing methods at the usual precinct granularity.

Wiggle: Physical Challenge-Response Verification of Vehicle Platooning

In an autonomous vehicle platoon, a set of vehicles coordinate with each other to travel single file with greatly reduced following distances. Platooning promises benefits such as fuel efficiency, road safety, reduced traffic congestion, and passenger comfort. Vehicle platoons are often envisioned as dynamic, with vehicles fluidly joining and leaving the platoon. This dynamicism, however, introduces a security risk as platoon members can send dangerous coordination messages to other vehicles. Intervehicular message integrity can be verified with cryptographic techniques, but message truthfulness cannot be, as an entity with the right keys could send a valid request to join a platoon, but not actually be in the correct physical location to join the platoon. In this project we developed a security protocol for vehicles requesting to join a platoon. The protocol essentially challenges the candidate vehicle to perform a set of longitudinal perturbations (i.e. back and forth wiggles), and a verifier vehicle at the back of the platoon records the measurements of the vehicle behind (if any), and accepts the candidate request if the recorded movements sufficiently match those in the challenge.

The project was undertaken as part of an NSF-funded REU at the University of Arizona, and I was primarily responsible for designing the wiggle protocol and analyzing its security.