Siddhant Agarwal

I am a third year Ph.D. student in the Computer Science Department at University of Texas at Austin, working with Prof. Amy Zhang and Prof. Peter Stone. I am interested in utilizing reward-free transitions for improving reinforcement learning be it through designing principled representation learning techniques or downstream algorithms independent of rewards. For this I have been looking at RL from the perspective of visitation distributions and utilizing unsupervised RL, goal-conditioned RL and distribution matching learning techniques. Previously, I completed my Dual Degree (B.Tech. + M.Tech.) in Computer Science and Engineering from the Indian Institute of Technology Kharagpur with a department rank 1. During my undergrad, I was a leading member of the Autonomous Ground Vehicles research group where I worked on a number of perception and planning projects for autonomous driving. I completed my bachelor thesis on "Reinforcement Explanation Learning" under the supervision of Prof. Abir Das.

• [Sept 2023]: Paper accepted at NeurIPS 2023.
• [Aug 2022]: Joined Ph.D. program in the Computer Science Department at the University of Texas at Austin
• [May 2022]: Completed my Dual Degree in Computer Science at IIT Kharagpur, securing rank 1 in the department
• [Oct 2021]: Paper accepted at NeurIPS 2021 workshop on Deep RL and Ecological Theory of RL.
  
• [Oct 2021]: Paper accepted at the NeurIPS 2021 workshop on eXplainable AI approaches for debugging and diagnosis.
  
• [July 2021]: Interned at Goldman Sachs
  
• [March 2021]: Paper accepted at the ICLR 2021 workshop on Security and Safety in Machine Learning systems.
  
• [Jan 2021]: Paper accepted at ICLR 2021.
  
• [Nov 2020]: Paper accepted and nominated for best paper at KR2ML workshop, NeurIPS 2020.
  

We introduce a framework for unsupervised RL to produce optimal policies for any language instruction.

We provide a representation that can produce optimal policies for any reward function in the environment.

The Robotics stack used by our RoboCup Standard Platform League that used RL to train high level abstracted robotics behavior.

A two stage simulator and real-world manipulator system designed as a air-hockey for to study effects of RL and IL.

We introduce a general framework for goal conditioned RL using f-divergences.

We introduce latent representations that can predict the behavior of the agent at future steps to improve generalization in RL.

We reformulate the process of generating saliency maps using perturbation based methods for black box models as a Markov Decsion Process and use RL to optimally search for the best saliency map, thereby reducing the inference time without hurting the performance.

We show that backdoored classifiers can be attacked by anyone rather than only the adversary. We propose an attack that generates alternate triggers for the poisoned classifiers.

We show that using reinforcement learning (or even simple heuristics) we can produce deceptively perturbed knowledge graphs that preserve the downstream performance of the kg-aumented models.

We use a hybrid CNN archticture that uses two image processing features to classify the images. The CNN architecture has significantly less number of parameters than any of the state of the art methods on GTSRB.

We worked to develop vision, planning and localization modules for a category 4 autonomous vehicle.

We investigate commonly used cluster management systems like SLURM and Condor. We further develop a fault-tolerant active-passive SLURM-like cluster management system.

We use parallization in a GPU to accelarate graph algorithms like BFS, DFS and Single Source Shortest Path and All Pair Shortest Path to achieve massive speedups.

We develop an android chatbot application using a trained LSTM based encoder-decoder module trained for emotion specific chats. We use a simple classifier to choose the appropriate model from the chat.

Webpage template courtesy: Jon Barron