INFORMS 2021 Program Book

INFORMS Anaheim 2021

WD09

Wednesday Keynote 04 CC Ballroom D /Virtual Theater 4 Keynote: Algorithms and Social Service Provisions Keynote Session 1 - Algorithms and Social Service Provisions Rediet Abebe, University of California-Berkeley, Berkeley, CA, 14853, United States Bio: Rediet Abebe is an Assistant Professor of Computer Science at the University of California, Berkeley and a Junior Fellow at the Harvard Society of Fellows. Abebe holds a Ph.D. in computer science from Cornell University and graduate degrees in mathematics from Harvard University and the University of Cambridge. Her research is broadly in algorithms and artificial intelligence, with a focus on equity and distributive justice concerns. As part of this research agenda, Abebe co-founded and co-organizes the MD4SG initiative and is serving as a Program Co-Chair for the inaugural ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ‘21). Her dissertation received the 2020 ACM SIGKDD Dissertation Award and an honorable mention for the ACM SIGEcom Dissertation Award for offering the foundations of this emerging research area. Abebe’s work has informed policy and practice at the National Institute of Health (NIH), the Ethiopian Ministry of Education, and the United Nations Food Systems Summit. Abebe also co-founded Black in AI, a non- profit organization tackling equity issues in AI. Her work is influenced by her upbringing in her hometown of Addis Ababa, Ethiopia. WD06 CC Room 303A In Person: Analytics for Migration and Resettlement General Session Chair: Buket Cilali, The University of Oklahoma, Norman, OK, 73069- 5703, United States 1 - Refugee Resettlement with a Time-evolving Minimum Cost Problem Approach Deniz Emre, University of Oklahoma, Norman, OK, United States, Buket Cilali, Kash Barker, Andres David Gonzalez The literature in refugee relocation focuses generally on short-term planning, but such problems are continuous and require long-term planning. Inspired by the long-term nature of the problem, we consider temporal networks, a special case of multi-layered networks in which time is incorporated with layers. With this snapshot representation of time for the resettlement problem such that each layer is a static representation of the network at a different point in time, we create a network of temporal layers. Our aim is to solve a minimum cost flow problem from the first to final layers while synchronously solving a resettlement facility location problem in each layer and managing the temporal opening or expansion of resettlement locations. 2 - Multi-stage Stochastic Programming for Long-term Refugee Resettlement Planning Buket Cilali, University of Oklahoma, Norman, OK, 73069-5703, United States, Kash Barker, Andres David Gonzalez Most studies in refugee resettlement assume that capacity/demand within a fixed time interval is given. However, refugee resettlement is not a one-time event. On the contrary, it is an ongoing and long-term process with dynamic parameters. To deal with the current conflict-based resettlement problem, as well as future climate-driven variants of the long-term displacement problem, we take a stochastic approach and adapt the supply change management framework. To this end, we address issues related to the nature of resettlement (e.g., cost of opening new locations to service, penalties/incentives for unmet demand, dynamic capacity/demand management, socio-cultural impact of the resettlement process). Wednesday, 2:45-4:15pm

WD07 CC Room 303B In Person: Probabilistic Forecasting/Operations of Knowledge-Intensive Services General Session Chair: Pavel Atanasov 1 - The Efficacy-revenue Trade Off In Pharma R&D and its Incentive Implications Stylianos Kavadias, Margaret Thatcher Professor of Innovation & Growth, University of Cambridge, Cambridge, United Kingdom, Jeremy Hutchison-Krupat, Konstantinos Stouras There exists a fundamental tension that underlies the R&D process for pharmaceutical drugs between achieving a sufficient efficacy level to advance a drug further and rapid development to ensure maximum opportunity to exploit the drug patent times. We explore differences in how this trade-off manifests itself amongst different stakeholders in the development process and how senior management can best adapt their incentives. 2 - Human Forest vs. Random Forest in Time-sensitive Covid-19 Clinical Trial Prediction Pavel Atanasov, Pytho LLC, Brooklyn, NY, United States, Regina Joseph, Felipe A. Feijoo, Max Marshall, Sauleh Ahmad Siddiqui What methods generate the most accurate forecasts about clinical trial phase success? We describe the first multi-method clinical trial forecasting tournament, comparing machine learning models and crowdsourcing methods that estimate the time-dependent probability of phase transition for COVID-19 vaccines and treatments. The crowdsourcing approach uses the Human Forest process and software, which enables forecasters to define custom reference classes, query a historical database and review resulting base rates. The base rates, and forecaster- adjusted probabilistic estimates, are aggregated. Accuracy was compared against a random survival forest machine model, across 28 questions. Results show that Human Forest significantly outperformed the RSF model, registering 32%-48% better Brier scores. Human Forest’s advantage was due to better calibration. WD09 CC Room 303D In Person: Optimistic/Robust Sequential Decision-Making General Session Chair: Rui Gao, University of Texas at Austin, Austin, TX, 78712-1277, United States 1 - Data-driven Optimistic Optimization and Contextual Decision-making Junyu Cao, The University of Texas at Austin, Austin, TX, 94706- 1991, United States, Rui Gao We study sequential decision making under uncertainty with contextual information, in which the decision-maker jointly learns a predictive model and optimizes a downstream decision task based on the relevant context. Inspired from the optimistic counterpart of data-driven robust optimization, we propose a novel framework that designs principled upper-confidence-bound algorithms with efficient implementation and strong performance guarantees. 2 - An Efficient UCB Algorithm for Contextual-bandit-based Learning with Continuous Actions Zhi Wang, University of Texas at Austin, Austin, TX, United States, Rui Gao In online learning and decision-making with contextual information, upper- confidence-bound (UCB) algorithms are a celebrated class of algorithms. Choosing an action with the highest upper confidence bound of the reward, each iteration of the algorithm involves a joint optimization over the action set and the parameter confidence set. When the action set is in continuum, this subproblem is computationally intractable, thus prevents the design of efficient UCB algorithms. In this paper, we propose an efficient UCB algorithm based on the first-order approximation of the optimal reward function. To the best of our knowledge, this is the first efficient UCB algorithm that achieves nearly optimal regret for a variety of problems, including linear bandits, generalized linear bandits, multi-product dynamic pricing, and parametric contextual bandits.

167