Active Search with Complex Actions and Rewards
Abstract
Active search studies algorithms that can find all positive examples in an unknown environment by collecting and learning from labels that are costly to obtain. They start with a pool of unlabeled data, act to design queries, and get rewarded by the number of positive examples found in a long-term horizon. Active search is connected to active learning, multi-armed bandits, and Bayesian optimization.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2017
- Accession Number
- AD1168010
Entities
People
- Yifei Ma
Organizations
- Carnegie Mellon University