Tweeted By @svlevine
Reward design in RL is a bit of a mystical art. We can start to demystify it using control as inference, and also learn rewards from data using variational inverse control with events (VICE): https://t.co/VmUeMLg4PN
— Sergey Levine (@svlevine) June 1, 2018
w/ Justin Fu, @avisingh599 Dibya Ghosh, Larry Yang