Turns out that simply extracting policies with supervised learning conditioned on the advantage can get state of the art results in RL, as long as we get a few important details right! https://t.co/fZsaoIcUgs