I am a third-year PhD student in computer science at the University of Pennsylvania. I am being advised by Rajeev Alur. I received my B.Sc. (Honors) in Mathematics and Computer Science from the Chennai Mathematical Institute.
My (broad) research areas of interest include formal methods, verification, programming languages and machine learning. In particular, I am interested in applying formal methods to improve applicability and reliability of deep reinforcement learning methods, verifying systems with machine learning components and using machine learning to improve scalability of program analysis and verification techniques.
Formal Methods for Reinforcement Learning. Reinforcement learning is a promising approach for learning control policies for robot tasks. This project focusses on applying techniques from logic and formal methods to enhance reinforcement learning. For example, how to compile formal logical specifications to reward functions so that one can use existing RL algorithms to learn policies for such specifications? How to use state abstractions for hierarchical reinforcement learning? How to verify/test learned policies for safety properties?
Streaming Algorithms over Probabilistic Streams. In streaming algorithms, streams are defined to be sequences of data points. In this project, we consider probabilistic streams which are sequences of distributions over a finite set of events. We analyze space complexity of streaming computation in this setting for different query classes.
Space-efficient query evaluation over probabilistic event streams, Rajeev Alur, Yu Chen, Kishor Jothimurugan, Sanjeev Khanna. In submission.
State and Action Abstractions for Hierarchical Deep Reinforcement Learning, Kishor Jothimurugan, Wenbo Zhang, Osbert Bastani, Rajeev Alur. In submission.
Techniques for verifying robustness of neural networks, Kishor Jothimurugan.
WPE II Survey Report.
PennCloud: Course Project for Software Systems (CIS 505). Built a cloud system with e-mail and storage services in a team of 4. I was responsible for the distributed backend key-value store based on Google BigTable that is sequentially consistent, fault tolerant and scalable.
Online learning with many experts: Course Project for Computational Learning Theory (CIS 625).
Survey of online learning methods for dealing with too many (potentially infinite) experts.
Testing non-interference with KLEE: Course Project for Software Analysis and Testing (CIS 700).
Used symbolic execution tool KLEE to test non-interference properties.