AI Poker Bot Pluribus Destroys Elite Players and Costs Less than $150 to Run

4 min read

Another artificial intelligence (AI) bot with “superhuman” powers has put the fate of poker players everywhere in jeopardy.

AI poker bot
AI poker bot Pluribus has shown that machines can beat high-level pros at their own game. (Image: CIO.com/Thinkstock)

Following two decades’ worth of development, AI poker software can now beat multiplayer No Limit Hold’em games.

Hailed by lead developer Tuomas Sandholm as a major breakthrough, Pluribus’ recent performance saw it defeat a collective of world-class players.

AI vs. Multiple Top Pros

Outlining the latest men vs. machines battle in Science Mag, Sandholm and Noam Brown put together two scenarios.

The first saw groups of five pros take on one version of the supercomputer. The other saw two pros take on a table of AI poker bots individually.

Among the 13 pros picked to play in the five-human/one-computer match was Greg Merson. With more than $11 million in live earnings and the 2012 WSOP Main Event to his name, Merson is a tough test for anyone.

However, after 10,000 hands and 12 days of action, not even the American’s skills could defeat Pluribus. According to the results, the AI software achieved an average win rate of 48 milli big blinds per game (mbb/game).

With one victory in the bag, Chris Ferguson aka Jesus and Darren Elias stepped into the breach. Each sitting down against five versions of the poker bot, the pros did what they could over the course of 10,000 hands.

Unfortunately for humankind, even the divine efforts of Ferguson couldn’t secure a win. Although the pair fared better, Pluribus still achieved an overall win rate of 32 mbb/game.

Five Frightening Facts: What We Learned from Men vs. Pluribus

  1. In 2017, Libratus was powered by 100 CPUs. In 2019, Pluribus required just two.
  2. Pluribus uses less than 128 GB of memory meaning it can run on a cloud server costing less than $150.
  3. Pluribus takes an average of 20 seconds per hand in a multiplayer setting or twice as fast as the average human player.
  4. Pluribus is completely self-taught. After programming in the rules of No Limit Texas Hold’em, the developers left it to learn and devise its own strategies.
  5. Although Pluribus has a blueprint for optimal play at the start of a game, it conducts real-time searches of its database and can adjust based on the moves of its opponents.

Self-Taught Poker Bot

Prior to taking on multiple players at once, the AI program played trillions of hands against itself. Using a technique known as “reinforcement learning,” it reached a world-class level in just 20 hours.

While that fact is impressive enough, Sandholm said the software’s ability to defeat multiple players at once was beyond expectations.

Although his previous project Libratus proved that one bot could beat one player, the task of defeating five top poker brains was significantly more difficult.

However, after 12 days of betting and bluffing, developers are now confident AI could not only thrive in poker but the real world.

In their opinion, being able to strategize against multiple opponents could help in areas such as cybersecurity, healthcare and finance.

Pluribus Provides Some Positives for Players

For poker players, the latest defeat is a potentially ominous sign. However, there are some positives.

From a strategy perspective, Pluribus’ tactics could revive the donk bet. Often maligned by experts, the act of calling a raise and betting into the raiser on the next street may be more effective than first thought.

Why the bot came to this conclusion is unclear. However, it made the move “far more often” during its matches than professionals.

Pluribus also reconfirmed that limping is a poor strategy. Although it initially experimented with the move during its learning phase, the program eventually determined that raising or folding were optimal.

While we may not be on the brink of an AI takeover, time could be running out. In 2017, research conducted by MatchPoker suggested that No Limit Hold’em was still too complex for AI to solve.

Fast-forward two years and Pluribus is proving that theory wrong. What’s more, the software operates on a cloud server that costs just $144 to run.

In summarizing the latest results, Sandholm and Brown admit that AI is far from perfect. However, even in complex imperfect-information settings, current technology can produce “superhuman” strategies.



Related Posts

Did you know about our poker forum? Discuss all the latest poker news in the CardsChat forum

Popular Stories