Speaker: Dhruva Kartik
Affiliation: University of Southern California
Abstract: Active hypothesis testing is a classical problem which naturally arises in various applications like anomaly detection, target tracking, communication with feedback and clinical trials. With the recent attention on autonomous systems and the Internet-of-Things, there is renewed interest in active hypothesis testing. Classical approaches for experiment design in active hypothesis testing tend to be fixed/open-loop and random. We design adaptive strategies and analyze the performance gains achieved in the non-asymptotic regime.
We consider two Neyman-Pearson type formulations. In these formulations, the agent can perform a fixed number of experiments and then decide on one of the hypotheses. For these problems, lower and upper bounds on the optimal misclassification probabilities are derived and these bounds are shown to be asymptotically tight. In contrast to open-loop, randomized strategies, our methods are fully deterministic and adaptive, while also asymptotically optimal. For a special class of anomaly detection problems, tighter non-asymptotic bounds are obtained. This tighter analysis suggests that our adaptive strategy is second-order optimal. This observation is reinforced by numerical experiments.
Biography: Dhruva Kartik received his B.Tech. degree in electronics and communication engineering from the Indian Institute of Technology, Guwahati, India, in 2015. He joined the Ming Hsieh Department of Electrical and Computer Engineering at the University of Southern California, Los Angeles, in 2015, where he is working towards his Ph.D. degree with Prof. Urbashi Mitra. His research interests are in decentralized stochastic control, decision-making in sensing and communication systems, game theory, and reinforcement learning.
Date(s) - Feb 13, 2020
1:00 pm - 2:30 pm
EE-IV Shannon Room #54-134
420 Westwood Plaza - 5th Flr., Los Angeles CA 90095