A lot has changed from late 2016. Perhaps the most significant is that AI can now clearly defeat humans in heads-up big bet poker games. We can thank the research invested into subgame solving techniques for this. While many public software tools use a simple form of unsafe subgame solving based on assumptions (and are usually still effective), there are “safer” alternatives that provide theoretical guarantees for reducing overall exploitability which are more powerful still.
For example, the University of Alberta combined safe resolving with a neural network that can estimate EV values in their DeepStack AI. This agent can naturally play heads-up no limit hold’em at any stack depth and intelligently respond to any bet size it faces. Next up, Carnegie Mellon used the reach max-margin technique in their Libratus agent, which defeated human pros in the 2nd edition of Brains vs AI. Noam Brown continued to find many more optimizations for resolving, summarized in his award-winning paper. Well deserved and thanks for the acknowledgement.
There was a 2017 computer poker AI contest, but it was less competitive since the latest techniques require more computation than the contest rules allow. I submitted a slightly modified version of Act1 called Intermission, which ended up winning both the instant runoff and bankroll categories.
Looking ahead, people wonder if multiplayer poker (with 3+ players) will be the next target. It may be closer to reality than many expect…