Historical milestone · 2017
Deep reinforcement learning from human preferences
Reviewed through September 18, 2026
2017 · Historical milestone
Deep reinforcement learning from human preferences
- Era
- 2010s
- Theme
- Safety, security & alignment
- Evidence form
- Human-in-the-loop experiments
- School / paradigm
- Preference learning / alignment
- Institution / context
- OpenAI; DeepMind
- Researchers
- Paul Christiano; collaborators
Understand
Plain-language record, transferred from the reviewed source module.
Theory or experimental setup. Learned reward functions from pairwise human comparisons and optimized agents on simulated control and Atari tasks.
Result / historical claim. Demonstrated that sparse preference feedback could train complex behavior without a hand-specified reward.
Apply
Professional implication, only where the reviewed record states one.
The checked-in record does not state a separate professional application for this entry. The topic page places it in the wider research lineage: Safety, security, and alignment.
Verify
Evidence status, stated limitations, and the external sources this record actually carries.
Evidence form. Human-in-the-loop experiments
Limitation / debate. Learned rewards can be incomplete or exploitable and depend on labeler consistency and coverage.
Source status. The source link below is the verified link our reviewed topic research already carries for this milestone.
Reproduce
A reproduction tutorial is linked only when one exists for this exact record.
A reproduction tutorial is not yet available for this entry. The closest reviewed material is Safety, security, and alignment.
Cite or share
APA-like: This historical record carries a year only, and no author or publisher of record in the checked-in data. An APA reference would have to invent that metadata.
BibTeX: BibTeX requires an author and publication venue. Historical lineage entries store a narrative record and its source link, not structured authorship, so the field would be fabricated.