Guide to Facebook Poker AI — Is Online Poker Dead?

Are we about to witness the end of online poker as we know it? Facebook has created a Poker AI named Pluribus that managed to beat 12 poker pros, playing against 6 of them at a time.

That’s something no other AI has managed to do before. Does this mean that the online poker market is never going to be the same again?

In this article, I am going to tell you everything you need to know about Pluribus and try to answer all the burning questions. Some of the topics I’ll cover include:

• What is Pluribus?
• What is Libratus?
• How does Pluribus work?
• What’s so special about it?
• The Nash Equilibrium and Blueprint strategies
• How Pluribus managed to beat 12 poker pros
• What are the practical uses of this AI?
• What does this mean for online poker? Is it dead?

With all these crucial questions in front of you, I’m sure you’re anxious to find out about the answers as well. Without further ado, let’s begin by learning what Pluribus actually is.

What Is the Pluribus Poker AI?

The Pluribus AI is Facebook’s newest achievement when it comes to artificial intelligence technology. This AI has managed to do something that has been unimaginable just several years ago — beat human opponents in six-player no-limit Texas Hold’em.

This AI is the result of the collaboration between Facebook and Carnegie Mellon University. Pluribus is basically an improved version of Libratus, its younger brother that achieved some incredible results in the past. However, beating elite players in a full-on poker game with several opponents was a milestone that had been missing for a long time.

This huge milestone is not important just for the poker industry. It matters for the entire world, as Pluribus can be used in different scenarios, like improving traffic conditions.

Moreover, Pluribus is special because it managed to reach this achievement all by itself. It is a self-play AI which does not require any input from humans in order to improve its skills, apart from the initial guidelines on the rules of the game it’s playing.

Libratus: Where It All Began

Libratus is the “old” Poker AI, also created by Carnegie Mellon University in 2017. At the time, it set a milestone by winning in a two-player Texas Hold’em game against human opponents, which was something that had not been done before. 

However, the Libratus’s computing limitations meant that it could only perform well in games with two opponents or two teams and zero-sum competition. Some games in which this AI used to excel include chess, checkers, Go, Dota 2, Starcraft 2 and two-player poker. 

To increase its chances of winning against human players, Libratus employed two strategies called the Nash Equilibrium and the Counterfactual Regret Minimization. These strategies are used in games with imperfect information in order to fill in the blanks. 

However, these strategies are very difficult to use and compute when there is more than one opponent in the game. The computing requirements increase exponentially as you add new opponents. 

For example, if Libratus would face off 5 human opponents, it would need 10,000x computing power that is needed for only one opponent. On the other hand, Pluribus needs just a fraction of that power at just a fraction of the cost, which makes it much more practical, efficient, and economical.

How Does Pluribus Work?

The scientists at Carnegie Mellon University came up with some incredibly clever solutions to make Pluribus much more capable and more intelligent than Libratus. 

The reason why Libratus needed so much more computing power was that it had to look much farther ahead.

Therefore, it had to do more calculations before making a move, sometimes predicting the moves and the outcome of the entire game. 

On the other hand, the secret behind Pluribus is that it needs to look just several steps ahead before making a move. This means that it requires minimal computing power, which even a standard CPU found in our homes can provide.

Unfortunately (and fortunately), Facebook did not disclose how Pluribus works in detail. We know that it uses a new, faster self-play algorithm to navigate imperfect-information games, but we do not yet know what it actually is. 

This is both good and bad. On one hand, we are in the dark and thus curious to find out what happens behind closed doors. On the other hand, fraudsters will not be able to recreate Pluribus without complete information from Facebook. At least for some time.

What Is So Special About Pluribus?

Pluribus is special and different from anything that we have seen before in so many different ways.

It’s an underestimation to say that it is a revolution in artificial intelligence research. 

Self-Play AI

First of all, Pluribus is a self-play AI. That means that it did not require any input from humans in the form of real-world examples from poker games. This also means that it did not learn from human-completed games. Instead, it taught itself. 

Pluribus received the guidelines regarding the rules of no-limit Texas Hold’em poker and then set off to play against itself until it reached perfection. Basically, it taught itself by making mistakes over and over again and figuring out what the perfect strategy would be. 

This might sound like a long process, but it is actually frighteningly fast. By training itself, Pluribus is able to exceed the average human performance in poker after just 400 minutes (~7 hours). It requires another 800 minutes to exceed top human performance and another 1400 minutes to eliminate limping from its performance.

Requires Little Memory

Since Pluribus does not require any input from humans, that means that its memory requirements are quite low. As a matter of fact, this initial achievement was reached with just 512 GB of RAM. Moreover, no GPUs were used in the process. 

This means that, if we consider the typical cost of cloud computing space, the development of Pluribus would cost around $150. That’s just a fraction of the money that was necessary to even think about building previous poker AIs, such as Libratus.

Incredibly Fast

When playing a game of poker, Pluribus runs on two CPUs. To put that into perspective, AlphaGO required 1,920 CPUs and 280 GPUs to perform real-time searches in 2016. 

On top of requiring less processing resources, Pluribus is also much faster than its predecessors. To complete a search on a single subgame, Pluribus needs between 1 and 33 seconds, depending on the individual situation.

On average, Pluribus plays twice as fast as the fastest human professional poker players.

Why Is Poker so Difficult for AI?

Poker is the most difficult game in the world for AI to master.

The most challenging one so far is the six-player no-limit Texas Hold’em. The three main reasons why this game is so difficult for artificial intelligence are:

• Having to deal with incomplete information.
• It’s challenging to achieve a Nash equilibrium.
• Psychological skills such as bluffing are necessary to win.

Imperfect Information

Poker is defined as an imperfect-information game because players never have a complete image of the game. There are some games that also include imperfect-information, but none of them to the same degree as poker. 

In poker, players know only which cards they have, but do not know their opponents’ cards. Therefore, they need to make decisions that are based on predictions and not on facts. That’s completely different from chess, for example, where you get all the information from the board. That allows you to solve the end of the game without knowing any previous moves and strategies.

In poker, the main challenge is that it is impossible to create the perfect strategy for a particular situation that is different from the overall strategy of poker.

Nash Equilibrium Challenges

The Nash equilibrium is a strategy that was developed some time ago by a mathematician called John Nash. This strategy is applicable in zero-sum games with two players and is quite difficult to achieve in games with multiple players such as poker.

The Nash equilibrium is a strategy that guarantees a win for the player, regardless of the moves his opponents make. For example, in the rock-paper-scissors game, the Nash equilibrium strategy advises randomly choosing rock, paper, or scissors, as all three moves have an equal probability of winning.

Libratus managed to use the Nash equilibrium strategy successfully in a two-player Texas Hold’em game. However, achieving this strategy in a six-player game is simply impossible from a computational point of view.

Human Psychology and Bluffing

Try as we might, we still haven’t built an AI that could successfully mimic the complex human psychology. In poker, human psychology can be best observed in bluffing ⁠— learning how to lie to your opponents and trick them that you have a better hand than you actually do. 

Before Pluribus was created, this was impossible to do on a large scale. The main objective when bluffing is to keep your bluffing strategy unpredictable. Therefore, you cannot bluff using simple, pre-learned rules. You need to vary your approach and constantly re-learn what you knew before.

Considering that Pluribus managed to beat 12 elite poker players, we can confidently say that it has managed to master human psychology, at least in the context of poker.

What Is Pluribus’s Blueprint Strategy?

The Blueprint strategy that the Pluribus AI uses is a coarse-grained strategy that has helped it to win against human elite poker players. Because of the complexity and size of six-player no-limit Texas Hold’em, a more fine-grain strategy would not work. There are simply far too many possible outcomes for each situation. 

Therefore, Pluribus improves upon the more coarse-grained Blueprint strategy in real time and refines it for each situation. That way, its gameplay remains unpredictable and it is able to better adapt to different challenges. 

Real-time searches are not new in AI research. Numerous AIs have used real-time searches in perfect-information games to win against human players. An AI would typically look ahead a certain number of moves until it reached a leaf node.

However, these search methods are not possible in imperfect-information games such as Texas Hold’em poker. That’s because AIs cannot predict with certainty what the players would do. Human players can shift their strategy at any time beyond the leaf node and exploit the AI’s stiff strategy.

Moreover, in poker, the perfect strategy cannot be created based solely on the rules of the game. Another important factor is which strategy the opponents are employing and what their perception of the game is. For example, if a player never bluffs, then we could confidently fold whenever they raise and our hand is weak because we can be certain that they have a strong hand. That’s a big challenge for artificial intelligence to figure out.

Pluribus Is Extremely Cheap to Make

Apart from the scientific and technological achievements Pluribus signifies, it also marks another huge breakthrough in AI research. Namely, this AI is incredibly cheap to make. That’s entirely opposite of what the situation was just several years ago when you needed hundreds of thousands or millions of dollars to achieve similar results. 

The reason Pluribus is so cheap is that it requires very little memory, as we have already mentioned. If you have the right knowledge and the time to write the code for Pluribus’s algorithm, all you need is around $150 to buy around 512 GB of cloud computing storage. 

This means that anyone who wants would have the means to make it. That is one of the main reasons why Facebook did not disclose all the information regarding the code and the algorithms, apart from the fact that they want to keep their technology to themselves. 

Moreover, if someone were to make Pluribus and use it to play poker in online casinos, their ROI would be huge very quickly. This AI is able to outplay the majority of average poker players after just 7 hours of self-play and isolated practice. 

Before Pluribus came around, the overall sentiment among AI experts was that the future of AI lies in the hands of massive teams of engineers with millions of dollars at their disposal. However, this achievement shows that anyone with the right knowledge and the right approach can push the limits of artificial intelligence as we know it today.

The Human Pro’s Stand No Chance Against Pluribus

Pluribus’s abilities when it comes to playing six-player no-limit Texas Hold’em were not tested against an average player. It played against the top professionals in the industry, three of which were Chris “Jesus” Ferguson, Greg Merson, and Darren Elias, all of whom have won major world poker championships several times. 

Moreover, all the players who faced off Pluribus have won more than $1 million by playing poker professionally, and many of them much more than $10 million. As you can see, the AI did not get an easy nut to crack. It was playing against some of the best players in the world at this moment.

Carnegie Mellon University

 

Even so, the human players stood no chance against the AI. During the tests, the games were played in two formats: one AI against 5 human players and 5 AIs against one human player. At the beginning of each game, each player started with 10,000 chips on a table with 50/100 big-small blinds.

At the end of the games, considering that each chip was worth a dollar, Pluribus would have won around $5 per hand and $1,000 per hour. In the world of professional poker, this is considered a decisive win. 

According to Darren Elias, Pluribus’s main strength was the ability to use different strategies to win. That is something human players often cannot do. Humans try to make their strategies as simple as possible in order to be able to make fewer mistakes. That’s something this AI does not need to worry about, so it’s able to shift its playing style frequently. 

Why Did Facebook Develop This Poker AI?

Even though this achievement of Facebook and scientists at Carnegie Mellon University was tested in poker games, the implications of this technology are much greater.

Most of the challenges that we need AIs to solve for us are incomplete-information situations. Therefore, this milestone is important as it will allow us to begin using artificial intelligence in many different real-world settings. 

Even though Pluribus is not perfect and there are many multi-agent settings in which this AI would not fare well, it is the most advanced AI we have in this sphere today. 

Some real-world scenarios in which this AI could give us a hand include frauds, cybersecurity, as well as dealing with harmful content. All these situations could be presented to the AI as incomplete-information settings, so it could use its strategies to come up with the best outcome. 

Pluribus could also be used in traffic control to predict traffic jams.

This way, road maintenance could be scheduled better, with minimal consequences to the quality of life of residents of a city. 

The applications are huge and are limited only to what we as humans can imagine. But, since Pluribus is not based on an open-source algorithm, it remains to be seen what Facebook will do with it.

What Risks Does Pluribus Pose for the Poker Industry?

It comes as no surprise that the latest achievements in artificial intelligence research caused a huge stir in the online poker industry. 

We have seen many experts questioning whether Pluribus could completely devastate the online poker market. 

One of the main concerns with this AI is that its code could end up in the wrong hands. Online casinos and online poker rooms already have to deal with poker bots on their websites. They are run by people who use them for fraudulent reasons in order to make money. 

Luckily, these AIs are still not as sophisticated as Pluribus. Therefore, website officials are able to recognise the patterns in their play and to sanction such accounts. 

However, the situation is quite different with the latest AI. Its strategy is much more sophisticated than anything the online poker scene has seen before. Therefore, if a fraudster opened several accounts and used a realistic winning approach, it would be very difficult to catch them. 

Last but not least, we should not forget that Pluribus can be made for just $150. Anyone with enough knowledge and average computing skills can make this AI and let it loose in an online poker room. 

Is Online Poker Living Its Last Days?

If Pluribus were to be released and made public, then yes, poker would be on its last legs. However, that is not happening for this very reason. That means that online poker is still safe, at least for now. 

But, in the near future, we can certainly expect sophisticated AIs to saturate the newer online poker rooms. The only way to prevent this from happening is for poker websites to equip with other AIs that are able to recognise a computer play. 

If computers like Pluribus sat at online poker tables, humans would simply stand no chance. Poker rooms would still be able to make a profit, so it won’t be such a big blow for them. However, the average poker player would have no place in such a situation.
 

Conclusion

The scientific community is incredibly happy about the latest achievements related to artificial intelligence research. However, the online poker community, even though fascinated, is becoming increasingly nervous. 

Pluribus still does not pose any risk to our favourite poker rooms, but that could certainly change in the near past. In that case, the days of the online poker industry as we know it would be numbered. 

What do you think about this AI? Do you think it poses a real threat to online poker? You can share your questions and opinions below.