Decision-Making with Information Discovery
Methods
We devise intelligent systems for robust adaptive decision-making under uncertainty, when part of the decisions decide on the information/data to be gathered over time (as in active learning). Our general solution methods apply to the Best Box problem, to Weitzman's Pandora's box problem, to R&D project portfolio optimization, and many more.
Social Good Applications
Robust Active Preference Elicitation to Mitigate Homelessness
Motivated from the problem of designing policies for allocating scarce housing resources that meet the needs of policy-makers at the Los Angeles Homeless Services Authority (LAHSA), we have devised novel models and algorithms for actively eliciting the preferences of policy- and decision-makers in an optimal fashion, when only a limited number of questions can be asked.
Preference elicitation at LAHSA: we propose to learn the policy-maker preferences over policy characteristics (e.g., fairness/efficiency trade-off) by asking them pairwise comparisons over policy outcomes or asking us to tell us how much they like a particular policy.
We have devised algorithms that strategically select, out of a potentially huge number of questions to ask the policy-makers, questions to ask to be able to offer them their preferred policy.
Design of Decision-Support Systems that Align with Human Value Judgements
Motivated by the problem of addressing ethical dilemmas faced by decision- and policy-makers during the COVID-19, we have devised a near-optimal and tractable solution scheme to ellicit and aggregate the (potentially conflicting and imprecisely known) preferences of stakeholders in a fair way.
Model of the healthcare system during the COVID-19 pandemic as a queuing system. Individuals with different characteristics and conditions arrive over time and are matched to resources that are available (e.g., CCU beds, ventilators). The resource allocation policy determines which patients get matched to what resource and in turn impacts the fairness-efficiency characteristics of the policy.
Preference elicitation for allocation of COVID-19 resources: we propose to learn the preferences of stakeholders over policy characteristics (e.g., fairness/efficiency trade-off) by asking them pairwise comparisons over policy outcomes.
Related Papers
Active preference elicitation via adjustable robust optimization
P. Vayanos, (*) D. McElfresh, (*) Y. Ye, J. Dickerson, E. Rice
Under review at Management Science, February 2020
Robust optimization with decision-dependent information discovery
P. Vayanos, A. Georghiou, (*) H. Yu
R&R at Management Science, September 2019.
Data-driven learning in dynamic pricing using adaptive optimization
D. Bertsimas and P. Vayanos
Working Paper, 2017.
Decision rules for information discovery in multi-stage stochastic programming
P. Vayanos, D. Kuhn, and B. Rustem
In Proceedings of the 50th IEEE Conference on Decision and Control, pp. 7368-7373, 2011.
Partner Organizations
Related Grants
CES Triage Tool Redesign and Implementation
Home for Good Foundation
Role: Co-PI (PI: Eric Rice)
Award ID: TBD
Total Award Period Covered: 01/01/2020-12/31/2022
Total Award Amount: $1,450,000
Own Share: $334,000
Designing Fair, Efficient, and Interpretable Policies for Allocating Scarce Resources
USC James H. Zumberge Faculty Research & Innovation Fund Diversity & Inclusion Grant Program
Role: Sole PI
Funded for the period: 07/2018-07/2019
Total award amount: $30,000
Own Share: $30,000
Socially Optimal Personalized Routing with Preference Learning
U.S. Department of Transportation, METRANS University Transportation Center and National Center for Sustainable Transportation
Role: PI (co-PI: Maged Dessouky)
Award ID: DTRT13-G-UTC57
Funded for the period: 08/2017-07/2018
Total award amount: $99,998
Own Share: $79,096