Give the proliferation of artificial intelligence technology in machines these days, it was only a matter of time before AI started to influence our everyday lives.
Even basic functions are performed courtesy of artificial intelligence, from hailing an Uber to finding a contractor on Google. Even the workplace is utilizing AI to improve our productivity.
Scientists continue to test the limits of the technology, and their most extraordinary test cases tend to come about in the most humble of ways. Recently, computer engineers at Carnegie Mellon University in Pittsburgh had their own eureka moment with a poker-playing robot that could change the way the game is played forever.
As you may know, poker is a game built on the premise of smart decision-making, despite various emotions and signals aiming to throw you off, from the bluff to the bad beat.
So what would happen if we were able to remove the emotional responses of poker, and instead just utilize an algorithm that played it hands based on statistical probability alone?
Then you’d have Pluribus, a piece of software recently let loose in an online game of Texas Hold’em against a host of professional players. The results were both astounding and a tad worrying…
The brains behind Pluribus are Noam Brown and Tuomas Sandholm, computer engineers who aim to test the limits of AI in everyday situations. The application of their algorithm to poker is the perfect synergy: they could show how ‘superior’ decision making, based on facts alone, could outperform us humble humans, who can make bad choices when there are too many emotion-led stimuli to consider.
And so Pluribus figuratively took his or her (it feels strange to refer to the software as ‘it’ or ‘them’!) seat at the poker table to test their skills against six other players.
What followed was a landmark moment for artificial intelligence, as the machine defeated all five of his/her opponents – the first time an algorithm has beaten more than two opponents in a game of poker simultaneously.
The experiment unfolded during a game of No Limit Texas Hold’em, which is often a staple in many online poker offerings, sitting alongside variations such as Omaha, SNAP and 7-Card Stud. Hold’em involves considerable challenges in terms of understanding the strength of your own hand, predicting the flop, the river and the turn cards, and all while trying to decipher whether your opponents are bluffing or not.
In layman’s terms, multiplayer poker has more variables for players to consider, which makes the feat of Pluribus even more impressive.
And the software wasn’t just playing his/her hand by the book. It would occasionally bluff too, which really does make the mind boggle as to the future potential of these ‘machines’.
Poker has often been used by scientists as a benchmark for AI in problem-solving and setting benchmarks for its evolution. The fact that artificial intelligence can now outstrip humans on the felt – and be taught to bluff, most disconcertingly – suggests the future of poker could be about to change forever.
Pluribus Sets Standard for Dynamic ‘Superhuman’ AI
What makes the achievements of the Carnegie Mellon University scientists all the more impressive is that poker is, by its very nature, a wholly unpredictable game, where reading opponents and situations are key to success.
You could quite easily, if you had the skills and the IQ, teach an algorithm to play perfect Texas Hold’em based on the cards alone by programming it with all the probabilistic information it needs to know when to call, raise and fold.
But to engineer Pluribus in such a way that it could bluff and also set traps for its fellow players…well, that is an extraordinary feat of technology.
It’s worth referring to advances in AI as far as other games, such as chess and checkers, are concerned.
We can trace artificial intelligence in chess back to the 1990s, when an algorithm built by IBM called ‘Deep Blue’ defeated the then world champion Garry Kasparov across a six-game contest.
The Russian had beaten Deep Blue in a series a year earlier, notably when the AI made a mistake that allowed Kasparov to claim the win, forcing IBM to go away and come back with a stronger, seemingly unbeatable version of their algorithm.
“I could see that it’s, you know, a one-way street,” Kasparov later told Business Insider. “That’s why I was preaching for collaboration with the machines, recognising that in the games’ environment, humans were doomed.”
Today, we even have machines beating machines at chess. Google’s DeepMind department created AlphaZero, an AI construct that mastered chess by playing millions of games against itself, discovering winning strategies and moves doomed to failure.
AlphaZero then took on Stockfish 8, generally considered the most advanced chess ‘engine’ in the world, in a 100-game series. Google’s bot didn’t lose a single encounter, winning 28 into the bargain.
The AI works by ‘searching’ its database of 60 million different positions every second, before eventually coming up with the ‘right’ one. Many of its moves look unfamiliar to humans, but as yet no mere mortal has been able to defeat AlphaZero.
Checkers had experienced its own superhuman AI moment three years prior to the day Kasparov was humbled on the chessboard.
The ‘victim’ this time was Marion Tinsley, the best checkers player in the world who reportedly hadn’t lost a single match in 40 years. He had his own man vs machine contest with Chinook, an AI programme developed by a computer engineer named Jonathan Schaeffer.
Tinsley and Chinook first battled it out in 1992, when the human was able to overcome the machine, so like IBM would a couple of years later Schaeffer went back to the drawing board and reconfigured his algorithm’s learning capacity once more.
They reconvened in 1994, with a series of matches played out over a two-week period. The first six all ended in a stalemate.
But then Tinsley, in a twist of ill timing, got sick, and he was forced to withdraw from the contest. The checkers master died seven months later, and so we’ll never know if he or Chinook would have gotten the upper hand.
A Question of Possibility
The reason that AI is able to play perfect chess or checkers is that there is only a finite range of moves that can be made.
Okay, so this is on an almost mind-blowing scale: there are more than 288 billion possible positions of a chessboard after just four moves. In checkers, there are 500 billion billion (yes, you read that correctly) possible combinations in a game.
But a super-smart piece of AI software could learn all of these, which explains why they are able to master their art.
In poker, the parameters are different and more complex, and that’s why we still await the emergence of the perfect machine. It is possible to teach AI how to play each hand based on simple probability, i.e. the statistical likelihood of winning based on the cards held.
But being able to ‘read’ opponents, configure what hand they have and whether they are bluffing, is something entirely different. And that’s why, for now at least, the poker community can sleep easy.