You might be interested to read https://deepmind.com/blog/article/Agent57-Outperforming-the-human-Atari-benchmark https://deepmind.com/blog/article/Agent57-Outp

You might be interested to read https://deepmind.com/blog/article/Agent57-Outperforming-the-human-Atari-benchmark https://deepmind.com/blog/article/Agent57-Outperforming-the-... which has an alternative approach that does better on those games. Basically those games involve lots of “exploration” to find the winning states, so you need some algorithm that is incentivized to explore through a large state space even when it hasn’t seen any reward there.