Refs

• TODO
  • https://awllee.github.io/smc-tutorial/smc-tutorial.html#42

Minimizing KL(q||p) = maximizing rewards (aka policy gradient) - a KL penalty

Twist functions