This was part of New Directions in Reinforcement Learning and Control

Success Conditioning as Policy Improvement: The Optimization Problem Solved by Imitating Success

Daniel Russo, Columbia University