#  Koulik Khamaru (Rutgers University) 

 



####  calendar\_today Date and Time 

 **February 10, 2025** 

 11:30AM - 12:30PM EST 

####  pin\_drop Location 

 **SEC 3.301/302/303**  



 

 [ Register Here arrow\_circle\_right ](https://tinyurl.com/CRCSSeminarKoulikKhamaru) 

 



 

## **Talk Title: Automated inference with the Upper Confidence Bound algorithm.** 

This talk examines statistical inference challenges in adaptive data collection. While adaptive sampling methods are increasingly prevalent in modern applications, they invalidate classical i.i.d.-based statistical tools. In this talk, we present the principle of stability, introduced by Lai and Wei (1982), which states that when adaptive policies "stabilize" asymptotically, one can perform uncertainty quantification using techniques designed for i.i.d. data even when sampling adaptively. We argue that the popular UCB algorithm satisfies such a notion of stability in both multi-armed and contextual bandit problems. In both cases, we prove a quantitative central limit theorem.

Our key technique is a deterministic characterization of the number of arm pulls for a UCB index algorithm and a new comparison principle between the UCB algorithm and its noiseless, continuous-time counterpart. We expect this new principle to be broadly applicable for general UCB index algorithms.

## Speaker: Koulik Khamaru

Koulik Khamaru is an Assistant Professor of Statistics at Rutgers University. He completed his Ph.D. in Statistics from the University of California, Berkeley under the supervision of Prof. Michael I. Jordan and Prof. Martin J. Wainwright. Prior to his doctoral studies, he received his Bachelor of Statistics (BSTAT) and Master of Statistics (MSTAT) degrees from the Indian Statistical Institute, Kolkata.

His research spans theoretical and applied aspects of statistics, machine learning, and optimization. His work focuses on the EM algorithm, statistical challenges in Gaussian mixture models, model mis-specification, factor analysis, and non-convex optimization. Recently, he has been particularly interested in efficient reinforcement learning algorithm design, with emphasis on uncertainty quantification and sequential decision making, as well as active learning and causal inference problems.



 

 



 

 

 Share on:- [     Facebook ](#)
- [     Twitter ](#)
- [     Linkedin ](#)
 


 Save: [ Add to calendar calendar\_today ](https://crcs.seas.harvard.edu/node/1850626/event-feed.ics)  Copy link link