Joint Colloquium
Dept. of Statistics & Department of Operations and Information Management, School of Business
University of Connecticut
Storrs, Connecticut
By
Michael N. Katehakis
Management Science and Information Systems
Rutgers University
http://www.rci.rutgers.edu/~mnk/
Multi-Armed Bandits: A survey of models and recent advances
ABSTRACT
After a survey of bandit models, their different formulations and their wide applications, we present recent work and solutions to two key open problems. The first deals with the situation in which a decision to engage a process is subject to a commitment of a, possibly stochastic, number or duration of activations before a change to a different process is possible. It is shown that these activation commitments are equivalent to a more general depreciation model for which optimal policies can be constructed using propitiously defined restart in state indices. The second problems deals with the model in which outcomes from different bandits are normally distributed with unknown means and unknown variances, for which the regret increase rate can be minimized by sequential 'upper confidence bounds' based policies.
The talk ends with presentation of recent work on analysis of rates of convergence of forcing policies and index policies.
DATE: Friday, February 13, 2015
TIME: 10:00 a.m.
PLACE: School of Business (BUSN) Room 211
Michael N. Katehakis is a Professor in the Management Science and Information Systems Department at Rutgers University. In addition to Rutgers he has taught at Columbia, Stanford, SUNY at Stony Brook and at the University of Athens, Greece.
He is noted for his work on Markov decision processes, Stochastic Models, data-driven analytics and their application to queuing, reliability and service systems.
His is a Fellow of INFORMS, an elected member of the International Statistical Institute (ISI) and a Senior Member of the Institute of Electrical and Electronics Engineers (IEEE).
For more information, contact: Tracy Burke at tracy.burke@uconn.edu