The Importance of Search in Poker

In order to get the raw neural net from like 3,000 ELO to 5,000 ELo, you would have to scale up the training and the neural net size by 100,000 X. That's just if you do it at test time. And we actually did a man machine poker competition in 2015. We thought it might be superhuman and play against the world's best players. But what we found was that adding search on the river increased the performance of your model by 100,00 X over the course of an hour.