Pluribus: what is the new poker bot capable of?
Given the complex structure of the information about Pluribus, and since there is already a lot of information available in English, we will not go deep in lots of texts about bot history, win rates, and principles of operation, instead we will briefly list its features, achievements, and more importantly, focus on our own interpretation of the possible impact of Pluribus in modern online poker.
What is the fundamental difference between Pluribus and all previous bots?
- Play 6-max. All previous bots, including the direct ancestor of Pluribus — Libratus (in 2017 it was the first bot to beat the Douglas Polk team-pro), only played heads-up. Their algorithms, based on the Nash equilibrium, weren't unable to consider all the variables of the game with a large number of players;
- Power consumption. According to the developers, Pluribus used a computer with 128GB of RAM and two 14-core processors during the game and a $150 cloud service. This is, of course, powerful hardware, but isn't a university supercomputer worth millions of dollars;
- A new algorithm for the selection of solutions. It was this improvement that allowed Pluribus to go beyond heads-up and reduce power consumption. This bot doesn't go through all the possible solutions in each case, but instead, based on the hands played already, it learns and immediately discards unnecessary solution trees and group those remaining according to each case.
Some sources added to this list the bot's self-learning ability (initially only poker rules were loaded into it), but Libratus was also based on the hands' database rather than pre-built solutions.
In what conditions and who did Pluribus beat?
The main reason to spread information about the bot was the results of its game against a group of 15 volunteers, many of them well-known poker players: Chris Ferguson, Nick Petrangelo, Greg Merson, Darren Elias, almost all of them MTT-players. Interestingly, the Facebook article also lists Linus "LLinusLLove" Loeliger, one of the world's strongest cash games players, but there isn't a word about him in the scientific material. So, you are in your right to choose who to believe more between the two of them: Facebook or science.
The game was played in the following format:
- One 6-max 50/100 table with a fixed stack of 10,000 chips, no rake, no ante, and very large timebanks;
- Three people played in the format of 5 bots + 1 human and showed at the end 10,000 hands an average win rate of 3,27bb/100;
- The rest of the players participated and an average of 4 hours in a 1 bot + 5 people format, and at the end, the bot won 4,77bb/100.
See below a 5-minute video of Pluribus vs. real opponents:
So, everyone can see how the bot plays and make a judgment about its strength.
Pluribus: real opportunities and impact on online poker
As a result, we got the following:
The new poker bot, Pluribus, in conditions that have never been seen (and never will not be in online poker), showed a positive winrate in cash games, without taking into account the rake, against a group of live tournament players (some of which would hardly beat NL50 in PokerStars today). By itself, that is a very good result for the Artificial Intelligence industry, that works in an environment with incomplete information and will be able to help society in several fields. But of course, we are interested in how the emergence of Pluribus will affect modern online poker. For better understanding this, we need to mention a few more facts:
- The algorithm of loss minimization (Monte Carlo Counterfactual Regret Minimization) used is well-known to almost all solvers;
- Before playing against real people, Pluribus played with itself on a server with a 64-core processor for 8 days; this means that the self-training of the bot, in theory, began with eight days of calculations in a solver with potent hardware, while its rivals most likely didn't know anything about GTO in cash games.
Pessimism in the poker community related to the emergence of Pluribus is based on the fact that all the information they got was that a self-learning 6-max bot with low power requirements was created, and soon new versions or copies can be bought for a hundred bucks and be used at any stakes using a home laptop. And as a result, modern online poker, like a game of humans vs. humans, will die, sharing, for example, the same fate of online chess.
The fears are well-founded, but don't take into account some facts:
- First, the Pluribus developers are not going to share their algorithms, and will not use them on online poker. That is, the development of such a bot by dishonest people can take more than one year;
- Secondly, even a bot similar to Pluribus is not trained to play at modern rooms, and such adaptation takes time and brings more problems because its undemanding hardware requirements are a little exaggerated as the initial training of such a bot still requires a very powerful and expensive PC;
- Thirdly, even if in five years there would be 100-level Pluribus, ready to play in real conditions at real rooms (with a rake, dynamic stacks, smaller timebanks, and at several tables), and available to be bought for hundreds or several thousand dollars and working in average PC's, it doesn't mean that the poker rooms will be idle during all this time while they see how the software that kills their business develops.
The last point is the main difference between the modern online poker industry and online chess, which ultimately lost the war to bots. Not only this item postpones the death of online poker, as the first two but also gives a chance to avoid such an outcome. We are not going to touch upon all possible options for changing poker rooms, which can significantly complicate or make it impossible to use bots, but there is no doubt that those measures will appear. After all, even those who think that minus or near-even bots can be beneficial for creating action and raking, cannot deny (with the exception for quick profit lovers) that bots like Pluribus are no longer needed at the tables.
Right now, the main real disadvantage for the online poker industry from the appearance of Pluribus is the incitement of the informational hype around it, which is subject to the thoughts of the non-poker press. Amateur players or those who just want to start playing, after reading the headlines about software that can beat professional players at a 6-max table, obviously diminish their desire to deposit their money into such industry.
Also, it's essential to understand that opponents of Pluribus were players who, in principle, don't play professional cash games online (only Ferguson's choice is worthy, and you know his history), and would hardly show positive win-rates at low-stakes on PokerStars. Also, the games took place in an ideal scenario without no rake, huge timebanks, and fixed stacks.
Fighting bots, even in their current state, is very important for the online poker's future. But no matter what "gifts" we get with the technological progress, if players and affiliates join their efforts in the fight against this disaster, the possible death of online poker will become a very, very distant reality.
Stay tuned on our Telegram channel for more EV+ news.