Research
I'm interested in reinforcement learning, machine learning, stochastic approximation, optimization, and deep learning. Much of my research is about novel RL algorithms that are optimal and sample efficient.
|
1. Dynamic Mirror Descent based Model Predictive Control for Accelerating Robot Learning
Soumya R Samineni*,Utkarsh Mishra*, P Goel, C Kunjeti, H Lodha, A Singh, A Sagi,
Shalabh Bhatnagar , Shishir Kolathaya
(*equal contribution)
International Conference on Robotics and Automation (ICRA), 2022
NIPS Deep RL Workshop, 2021 (Poster)
NIPS Offline RL Workshop, 2021 (Poster)
project page
/
arXiv
/
video
Summary: Dynamic Mirror Descent is applied for an H step lookahead policy optimisation to augment the dataset for training an offpolicy RL, improving significantly the sample efficincy of Soft Actor Critic, widely used offpolicy RL algorithm. Further the proposed framework, DeMoRL generalises existing Model Based-Model Free (Mb-Mf) Approaches and acheives state of the art performance in Benchmark MuJoCo Tasks.
|
2. Policy Search using Dynamic Mirror Descent MPC for Model Free Off Policy RL
Soumya R Samineni*,
Masters Thesis, 2021
arXiv
/
video
/
code
|
|