Research

I am broadly interested in decision-making under uncertainty. My theoretical research focuses mainly on information collection problems, where decision-makers observe the outcomes of past decisions and use this knowledge to guide future ones. My applied research focuses mainly on problems of public interest, related to humanitarian logistics, medicine, and equity. Please see my complete list of publications for a more comprehensive overview.

Major areas of application

In recent years, my applied work has taken on the themes of public interest and social good, manifested in a variety of problems: 

• Non-profit management. In the paper “Cultivating disaster donors using data analytics” (Management Science, 2016), I studied a massive dataset of over 8.6 million direct-mail communications with American Red Cross donors, and identified design elements that positively impacted response rates.

• Medicine. My recent work, published in a medical journal (Drugs, 2021), used deep learning to study personalized multimorbidity management for type 2 diabetes patients. I found that clinicians make near-optimal decisions when treating a single disease, but there is potential for AI to do better when multiple diseases ("multimorbidities") are present.

• Humanitarian logistics. Recently, I was awarded a multi-year NSF grant (CMMI-2112828) to study post-earthquake reconnaissance planning. In the immediate aftermath of a disaster, public officials must rapidly assess the damage throughout a region based on data collected from a limited number of field inspections, which must be chosen optimally.

• Diversity, equity, and inclusion. My recent paper, accepted by Production and Operations Management, shows that ignoring sensitive attributes of job applicants produces outcomes that are systematically correlated with those attributes – and, paradoxically, this correlation can be removed by a policy that considers those attributes explicitly.

Algorithmic research

Much of my work, from the beginning of my career, has dealt with decision problems involving information collection. In these problems, we make choices based on uncertain predictions and observe their outcomes in the field; as a result, new data are obtained, changing subsequent predictions and decisions. Each decision thus serves two purposes: it has immediate economic value, and it also acquires new information that can improve future decisions. This dual role of decisions, which is often called "exploration vs. exploitation," frequently arises in healthcare, marketing, pricing, hiring, and many other domains.

Early on in my career, I studied the "value of information" approach as a way of achieving these goals. This methodology uses Bayesian statistics to evaluate the expected impact of an experiment ("expected improvement" or EI) on the economic objective of the problem. This quantity is measured in the same units as the objective (costs, earnings, health outcomes, clickthrough rates), creating an explicit tradeoff between the economic value of a decision and the value of the information that will be learned as a byproduct of it. The following papers are representative of this stream of work:

• Ryzhov, I.O. & Powell, W.B. (2011) "Information collection on a graph." Operations Research 59(1), 188-201. (PDF)

• Han, B., Ryzhov, I.O. & Defourny, B. (2016) "Optimal learning in linear regression with combinatorial feature selection." INFORMS Journal on Computing 28(4), 721-735. (PDF)

• Ryzhov, I.O., Mes, M.R.K., Powell, W.B. & van den Berg, G.A. (2019) "Bayesian exploration for approximate dynamic programming." Operations Research 67(1), 198-214. (PDF)

A highlight of this stage of my career was my single-authored paper "On the convergence rates of expected improvement methods" (Operations Research, 2016), which was the first to provide a theoretical explanation for why expected improvement has often been observed to work well. The reason is that, over time, it allocates the budget in a way that closely approximates a certain theoretical optimum derived in a totally separate stream of research. This discovery led me in a new direction focusing on the study of optimal allocations of the learning budget:

• Chen, Y. & Ryzhov, I.O. (2019) "Complete expected improvement converges to an optimal budget allocation." Advances in Applied Probability 51(1), 209-235. (PDF)

• Chen, Y. & Ryzhov, I.O. (2023) "Balancing optimal large deviations in sequential selection." Management Science 69(6), 3457-3473. (PDF)

• Zhou, J. & Ryzhov, I.O. (2023) "A new rate-optimal sampling allocation for linear belief models." To appear in Operations Research. (PDF)

In short, I developed new algorithmic concepts for learning optimal allocations of experimental budgets. Each paper focuses on a specific problem class, but the ideas are more broadly applicable and provide blueprints for designing methods of the same type in other problems.

Statistical research

Over time, I began to find that the biggest and richest challenge in decision-making under uncertainty was often not optimization per se, but rather, statistical modeling. To put it differently, these problems require principled statistical models that provide rigorous guarantees (coverage or accuracy), but are also able to interface with optimization methods in a way that enables tractable computation. Such issues were the focus of several papers:

• Qu, H., Ryzhov, I.O., Fu, M.C. & Ding, Z. (2015) "Sequential selection with unknown correlation structures." Operations Research 63(4), 931-948. (PDF)

• Chen, Y. & Ryzhov, I.O. (2020) "Consistency analysis of Bayesian learning under approximate Bayesian inference." Operations Research 68(1), 295-307. (PDF)

• Chen, Y., Marković, Ryzhov, I.O. & Schonfeld, P. (2022) "Data-driven robust resource allocation with monotonic cost functions." Operations Research 70(1), 73-94. (PDF)

My ongoing research has a prominent statistical component as well. For example, my NSF-funded work on post-earthquake disaster inspection studies a routing problem where vehicles are tasked with collecting data (rather than serving demand or acquiring rewards), and the objective function measures the quality of the dataset using an experimental design criterion.