Expanding the Frontiers of Decision Making Under Uncertainty
Abstract
The proposed research concerns some of the basic questions that arise when machines make decisions in the presence of uncertainty. Algorithms for these problems must make decisions that hedge against the unknown data. Many natural problems of this form feature information that arrives gradually over time, and irrevocable decisions that must be made along the way. Here, the decision-maker s algorithm must hedge against an uncertain future. There are also settings where information must be actively acquired prior to decision making. Acquiring information can be costly, time consuming, or dangerous, necessitating decisions on which information is most pertinent. In other settings still, the collection of information may need to be delegated to one or more agents, constraining how and which information can be collected and used. The proposed research will push the frontiers of automated decision-making under uncertainty in the following three directions- (1) Modeling correlation in data, and making decisions that hedge against such correlation. (2) Identifying a sparse subset of the data which suffices for making near-optimal decisions. (3) Delegating information gathering and decision-making to self-interested agents. Potential Impact- Understanding the impact of correlation, and the development of algorithmic techniques for hedging against it, can automate more robust and reliable decisions in partially-specified environments with hidden variables. Sparsification techniques for identifying the most consequential pieces of information are paramount when decisions must be made quickly, or when collection of information is costly or dangerous. Finally, protocols that algorithmically delegate information gathering and decision making to teams of agents, while incentivizing them to exert effort and faithfully communicate findings, can enable distributed and collaborative problem solving involving machines and humans.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2025
- Source ID
- FA95502410261
Entities
People
- Shaddin Dughmi
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Southern California