学术报告第201978期

题目：Learning from Data to Select Optimal Decisions and Policies

报告人：Yannis Paschalidis（波士顿大学）

报告时间：2019年11月27日 9:30--10:30

报告地点：南一楼 中 311

报告摘要：

Making optimal decisions in the absence of an accurate model predicting their impact

is a hard problem, particularly for sequential decision making under

uncertainty. Essentially one is faced with the double challenge of not knowing a

transition model and having to deal with the curse-of-dimensionality. I will consider two related settings. In the first, simpler setting, we are interested in making a single decision. A motivating example is to select the type of treatment to be recommended for a specific patient. I will present a new prediction-based prescriptive model that (i) predicts the outcome under each action using a robust nonlinear model, and (ii) adopts a randomized prescriptive policy determined by the predicted outcomes. The predictive model combines a new distributionally robust learning approach with K-Nearest Neighbors (K-NN) regression,which helps to capture potential nonlinearities embedded in the data. We establish out-of-sample guarantees for the predictive model and the optimality of the randomized policy in terms of the expected true future outcome. I will discuss applications in making treatment recommendations for patients with hypertension and for diabetic patients.

In the second setting, I will consider a Markov Decision Process and develop a robust

method for learning the policy and transition probability model from data (state,

action, next state tuples). We propose two robust maximum likelihood estimation

algorithms for learning the transition probability model and policy, respectively. An

upper bound is established on the regret, which is the difference between the average

reward of the estimated policy under the estimated transition probabilities and that

of the original unknown policy under the true (unknown) transition probabilities. We

provide a sample complexity result showing that we can achieve a low regret with a

relatively small amount of training samples. I will discuss applications in making

treatment decisions to optimize disease progression and also in robotics.

报告人简介：

Yannis Paschalidis is a Professor and Data Science Fellow in Electrical and Computer

Engineering, Systems Engineering, and Biomedical Engineering at Boston University. He is the Director of the Center for Information and Systems Engineering (CISE). He

obtained a Diploma (1991) from the National Technical University of Athens, Greece,

and an M.S. (1993) and a Ph.D. (1996) from the Massachusetts Institute of Technology (MIT), all in Electrical Engineering and Computer Science. He has been at Boston University since 1996. His current research interests lie in the fields of systems and control, networks, optimization, operations research, computational biology, and medical informatics.

Prof. Paschalidis' work has been recognized with a CAREER award (2000) from the

U.S. National Science Foundation, the second prize in the 1997 George E. Nicholson

paper competition by INFORMS, the best student paper award at the 9th Intl. Symposium of Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt 2011) won by one of his Ph.D. students for a joint paper, an IBM/IEEE Smarter Planet Challenge Award, and a finalist best paper award at the IEEE International Conference on Robotics and Automation (ICRA). His work on protein docking (with his collaborators) has been recognized for best performance in modeling selected protein-protein complexes against 64 other predictor groups (2009 Protein Interaction Evaluation Meeting). His recent work on health informatics won an IEEE Computer Society Crowd Sourcing Prize and a best paper award by the International Medical Informatics Associations (IMIA). He was an invited participant at the 2002 Frontiers of Engineering Symposium organized by the National Academy of Engineering, and at the 2014 National Academies Keck Futures Initiative (NAFKI) Conference. Prof. Paschalidis is a Fellow of the IEEE and the founding Editor-in-Chief of the IEEE Transactions on Control of Network Systems.