London AI developers Deepmind have been trying to “solve intelligence” since they were founded in 2010. Acquired by Google in 2014, Deepmind has focused on creating a neural network that can learn how to play strategic board games and video games, while also accessing external memory in a way the replicates the ways humans use short term memory.
The idea of Deepmind’s project is to combine “the best techniques from machine learning and systems neuroscience to build powerful general-purpose learning algorithms”, in an effort to not only successfully endow intelligence to machines, but also further the understanding of how the human brain works.
Using games for this purpose has proven to be the most successful way to create this understanding, as it allows the Deepmind to play the game and learn from its mistakes until it can come up with an optimal gameplay method. Perhaps the most interesting part is that, unlike IBM’s Deep Blue or Watson, Deepmind isn’t preprogrammed for just one particular game, but rather develops an understanding of strategy through experience whose principals it can potentially transfer to other games as well.
With AlphaGo, Deepmind’s latest and most famous AI program, being retired last month, is it possible that they can try to tackle other games that have so far been unbeatable by machines like Blackjack and Poker?
While Deepmind was first deployed on 70s and 80s video games like Pong, Breakout and Space Invaders, and quickly learn to play them beyond the ability of any human, in 2015 AlphaGo was introduced to the world.
AlphaGo’s purpose was to tackle one of the most difficult games for AI to win – the ancient Chinese abstract board game Go. While Chess offers players a branching factor of 35, meaning that on average a player has about 35 different legal moves that he can contemplate, Go offers a whopping 250, making it a lot harder for AI to correctly assess the optimal next move.
After playing a large amount of games that allowed it to evolve its algorithm, eventually AlphaGo got to the point where it could handily beat any other Go programs, and it was decided it was time to test it against human players. In October 2015, AlphaGo managed to defeat the European Champion Fan Hui, 5 to 0, in the first time ever that a computer program had beaten a professional human player without a handicap. But it wasn’t till this past May that AlphaGo took on Ke Jie, the world’s #1 ranked player and managed to beat him 3-0, thereby declaring the game Go as solved and retiring into the sunset, but not before leaving behind 50 games of Alpha Go vs Alpha Go as a gift to the Go community.
In 2015, Carnegie Mellon University’s Claudico poker bot played against four of the best poker players in the world and performed quite admirably – albeit ending the match over 700,000 virtual dollars down. Cepheus, another poker program developed by Michael Bowling at the University of Alberta in Canada, managed to turn a small win rate in the game of Limit Hold’em- which of course possesses a lot less variables and complex decision-making than its more famous cousin No Limit Hold’em.
When it comes to No Limit Hold’em, AI runs into a problem that it doesn’t really encounter in games like chess, Go, or Limit Hold’em: The ability to bluff incorporates a psychological aspect of the game along with the mathematical. This prevents the programs from simply relying on a purely statistics-based betting system, and would instead require it to attempt to decipher not only what the other players hand may be, but also see past any misdirection his opponents might engage in and understand their true intentions. While this may indeed come to be a possibility in the future, it is not currently possible with today’s technology.
Blackjack meanwhile is also a game that has yet to be fully solved. While it is certainly a much simpler game than poker in that there is an optimal way of playing the game based on the number of decks it is being played with, if you are playing with other players against the house you will always have to take into consideration how the decisions of other players will ultimately affect the dealer’s hand. Not every player in the table will be playing with the same optimal strategy as you, and that might mean that even if you play your hand perfectly, other player’s gameplay will ultimately lead the house to a hand that will beat yours.
Therefore, a professional Blackjack player often has to rely on his intuition and feel for the situation rather than simply stick to a chart. He will take into consideration how the table is playing and whether too many high cards or low cards have come in a row, to make his decisions, relying on endless hours of playing experience to know when to throw out the book and go with his gut. If you want to practice your Blackjack skills and gain such experience, Ninja Casino offers aspiring blackjack professionals the chance to practice for free.
Ultimately, since it’s almost impossible for everyone in the table to be playing at an optimal strategy, it simply cannot be said the game of blackjack has been truly solved. And while Poker and Blackjack both rely heavily on mathematics, Deepmind’s AI will have to seriously upgrade its neural network if it is to manage to incorporate the psychological aspects of each game in order to truly compete with the best players in the world. Google may have reached a new landmark in solving the ancient Chinese game of Go, but it still has a long way to go if it ever wants to solve the mysteries of the human mind.