The battle between man and machine in the gaming arena stretches back to the 1950s, and it’s one that still rages on today. In March 2016, the world’s leading Go player, Lee Sedol, from South Korea, was beaten by AlphaGo, an artificial intelligence (AI) program from Google’s DeepMind department. The loss was not only a shock for Go players around the world, but it was another sign that computers are now getting better at games.
While Sedol probably never anticipated that a computer would be his toughest test, there are some who believe computers will eventually be able to beat humans in any game of skill and logic.
Computer programmers were faced with a problem when they were first trying to create a machine that was capable of beating the top chess players. They needed to make their computer powerful enough to solve an incredible amount of possibilities per round using systematic analysis.
Because computers couldn’t use intuition to assess the game and think five moves ahead (the standard for any top player), it had to process each of the possible 30 moves for every chess position (this is an average) and repeat this process for up to five moves into the future. This problem stunted the growth of chess computers for many years until the advent of alpha-beta pruning in 1957.
A Movement Towards Solved Games
Alpha-beta pruning trimmed down the number of processes a computer would have to evaluate by implementing a system whereby the evaluation is stopped if a move proves to be worse than the previously examined move.
This breakthrough, combined with an increase in computing power, led to huge advancements in chess computer technology. By 1996 (11 years after it was first developed), Deep Blue was able to beat Garry Kasparov in a single match. Thanks to years of research, Deep Blue was able to evaluate move “goodness” (a technical term to describe how successful a move is likely to be) at a rate of 200,000,000 moves per second. These live calculations allowed it to think more than 10 moves ahead and, eventually, beat Kasparov.
By 1996 (11 years after it was first developed), Deep Blue was able to beat Garry Kasparov in a single match. (Image Credit: idisrupted.com)
From that point on, chess has been classed as a “partially solved” game by computer programmers and mathematicians. Unlike the common definition of “solved,” computer programmers use the term to describe a game where the outcome (win, lose or draw) can be correctly predicted from any position if both players are playing perfectly.
Essentially, if the computer is able to run through every possible move, assuming both players are playing an optimal strategy and select the one with the greatest expectation, then the game is considered solved. Today, a variety of games, including checkers, Connect Four and tic-tac-toe, are all solved while chess is only partially solved (some positions have been solved but not all).
No X Factor for AI
While computers might have the edge in chess and Go, they currently don’t hold the edge in blackjack and poker. Although it’s possible to point to Carnegie Mellon University’s Claudico poker bot (created in 2014 as the second version of the program Tartanian) and say it performed well in its match against four top pros in 2015, the fact remains that it didn’t dominate. In fact, it actually lost by more than 700,000 virtual dollars. One fact that proves the hypothesis that some games are simply too complex for computers to crack (for the time being at least), is that Cepheus, the poker program developed by Michael Bowling at the University of Alberta in Canada, could only play Limit Heads-Up poker.
Carnegie Mellon’s Claudico poker bot lost a heads-up No Limit Hold’em match in 2015 by 700,000 virtual dollars. (Image Credit: strategy.minnim.org)
Although Bowling wrote in the publication Science in 2015 that Cepheus had “weakly solved” Limit Hold’em and had an expected win rate 0.000986 big blinds per game (a very small overall win rate), the program couldn’t deal with No Limit Hold’em where the variables in each decision are much more complex.
Because the betting is limited in Limit Hold’em (hence the name), the process of bluffing, value betting and re-raising is extremely restricted and, therefore, players are forced to play in a much more mechanical way in Limit Heads-up Hold’em. This dynamic is great for AI machines because they can simply run various calculations and come out with the right move, much like they can in chess.
However, when you look at No Limit Texas Hold’em where the betting isn’t restricted, you find that there’s much more scope for diverting away from a purely statistics-based betting system. Indeed, when it comes to bluffing there is almost as much psychology involved in making the right move as there is mathematics. That’s something computers simply can’t handle yet.
Essentially, the major issue for AI computers is the aforementioned matter of intuition. Indeed, when you breakdown the personality traits of successful blackjack and poker players, you start to see that an ability to think around a problem and consider other variables is crucial in these games.
A professional blackjack player’s personality is a good example. Examinations of the traits of blackjack players often find them to be thoughtful, rational and capable of making optimal decisions – sounds like a computer, right? Still, it’s true that some blackjack moves have a greater mathematical expectation than others, but it’s also true that this may not always make them the best move in the context of the game.
For example, the math of the game suggests that hitting on 16 when the dealer is showing a nine is the correct move. However, if the player has previously seen six low cards dealt from the deck, their intuition would say that the book may be wrong on this occasion as a high card is coming.
Although this point could be debated in game theory terms (i.e. the move is optimal regardless), the fact remains that there aren’t defined “perfect” moves in blackjack and that makes it hard for a computer to compete against a human. Just like poker moves and bets.
As the definition of “solved” states, a game is only solved if all outcomes can be predicted from any position when both players are playing perfectly. If there’s no way to tell if the opponent is playing perfectly (because there are no “perfect” moves), then it’s hard – if not impossible – for a computer to solve the game.
We Aren’t a Beaten Race Just Yet
Similarly, the main characteristics of a top poker player are a combination of mathematical expertise, psychology, emotion and gut instinct. While a computer can excel on a mathematical level, it has no way to process instinct, psychology or emotion and that’s why AI often fails in games of No Limit Texas Hold’em.
Because many moves, such as bluffing, are based on the previous moves of an opponent, an insight into their emotional state and your own image at the table, it is quite difficult for a computer to compete. While they can come up with a solid solution based on the pot odds and the odds of a player holding a particular hand, they count account for that X factor.
Blackjack and poker are still dominated by humans and it doesn’t look like this dynamic is going to change any time soon. (Image Credit: allcars.pw)
For the likes of Tesla founder Elon Musk, the rise of AI is worrying in some respects as it could lead to computers educating themselves at a rapid rate and eventually having the ability to think on their own. After founding Tesla Motors in 2003 and creating a company with assets worth more than $7 billion, Musk began to review the growth of AI and became concerned with what he saw.
Bill Gates, who founded Microsoft in 1975, is also of the opinion that AI development needs to be monitored closely. However, there are always two sides to every debate. Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence (founded in 2013 by Microsoft co-founder Paul Allen), has said that we need to separate science from science fiction and realize that AI is a long way from perfection.
While he cited drones’ inability to pick and deliver shoes, it’s also true that AI still doesn’t dominate the games world. As we’ve shown, blackjack and poker are still dominated by humans and it doesn’t look like this dynamic is going to change any time soon.
There’s no doubt that in games such as chess and Go where the parameters for the “perfect” move are well-defined, computers can win. However, when it comes to poker and blackjack where there are intangibles involved in every decision, computers come up short time and time again.
So, are we a beaten race? The answer, for now, has to be a resounding no.