DAgger

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2024 Apr 4 13:56
Editor
Edited
Edited
2024 Jun 7 3:38
Refs
Refs

Expert policy가 필요한 단점

Compounding Error
에러에 도움은 된다 to address
Distribution Shift
  • Roll out learned policy
  • Query expert action at visited states
  • Aggregate corrections with existing data
  • Update policy
 
 

human gated DAgger

Expert intervenes at time when policy makes mistake
 
 
 
 
 

Recommendations