The best way to describe the recent rise of artificial intelligence in poker is that it’s a slow, methodical process. In fact, it took the most advanced poker playing programs on Earth almost four years to conquer the game and become the first AI to officially win at a major tournament. It started with a small group of programmers working on a simple program that could beat other AIs at online games like chess and Go. Once they had some success, they moved on to more complex games like Connect Four, but even then their progress was slow. The reason for this isn’t just because it takes a lot of computing power — although that’s certainly part of it — it’s also because winning at these games requires a huge amount of trial and error.
Poker is different. It doesn’t have any rules that you can memorize or patterns you can learn; instead, each player has an entirely unique set of skills that no other player possesses. You might know how to read your opponent’s tells, but unless you can anticipate what he’ll do, these signals will be useless. But if you can figure out which cards he’s holding, you can make educated guesses about his next move, based on what he’s already done. This is where AIs excel and why they’re so good at poker.
It turns out that the key to beating top-notch human players is to learn all that information from as many matches as possible and find ways to use it to your advantage. That’s exactly what AlphaZero achieved by using deep reinforcement learning. Deep reinforcement learning is a type of machine learning that uses neural networks to try to solve problems while improving over time. In this case, it used the same basic approach as AlphaGo Zero, the previous version of AlphaGo developed by Google subsidiary DeepMind. Instead of having a specific goal, such as moving to the right tile on a board, AlphaZero simply tried to win every match against its opponents.
In a sense, AlphaZero was programmed to play itself hundreds upon thousands of times until it realized that it could beat humans at the game without being told how to play. Then, when faced with real opponents, it would take into account everything it learned during training to maximize its chances of victory.
The person can visit on the OK Betting and either play the casino game or place the bet on their favorite sport. They can make good money in the long run through this option. A person need be really careful in choosing the best platform available for the people.
That’s not the only difference between the two AIs though. While AlphaGo Zero relied on Go-like features such as position, AlphaZero didn’t use any of them. Instead, it focused on every single card in every single hand it played. And because it couldn’t rely on visual cues, it was forced to make more educated guesses than before. All of this made it considerably better than the Go-playing AIs that preceded it.
But even after it reached this level of skill, it wasn’t done yet. At the beginning of 2019, DeepMind released a new version of AlphaZero called AlphaZero 2.0, which included several improvements meant to help it deal with the unpredictability of actual human opponents. For example, it now had access to real-time data showing how players were acting in the course of a match. After analyzing that data, it would adjust its strategy accordingly.
Even with these changes, however, it still wasn’t able to reach full mastery of the game. That changed last year, when DeepMind decided to test it against another A.I., one that had been trained using a similar approach. The company had built a computer code named Libratus, which could also beat top professional poker players. It had won six World Series of Poker (WSOP) tournaments since 2011, including three in 2018. DeepMind pitted the two A.I.’s against each other in a series of 20,000 hands of Heads Up No Limit Hold’em, a variation of Texas Hold ‘Em where there are no restrictions on betting.
The results were astounding. Both A.I.’s performed equally well, with the exception of a few occasions when Libratus held a slight edge. On average, both computers were able to win 50 percent of their matches while making around $100,000 per hand. When it came to money, there was no clear winner.
These results were impressive enough that DeepMind decided to release Libratus as open source software to the public, hoping that others would study it and improve on it. Since then, several groups around the world have launched projects aimed at creating their own versions of Libratus, including one team from Carnegie Mellon University.
So much for poker. What about other games? Well, the same techniques that worked so well for poker are also being applied to Chess and Bridge. One team from DeepMind recently managed to create an AI that could defeat human grandmasters in 10 minutes matches. Now that’s something we can get excited about.