Chess has for a number of years been seen as one of the ultimate games of skill and complexity, which is perhaps why it has had such an important role to play in the development of artificial intelligence.
The idea of a machine being able to beat a human at the game became a point of fascination for a great number of years, right up until that moment in 1997 when it was confirmed that IBM’s famous Deep Blue computer had beaten the then-reigning world champion Garry Kasparov in a match. Incredibly though, a recent development is now being hailed as perhaps a more important moment in AI than that event.
An AI with style
At the end of last year, the AlphaZero chess-playing computer developed by Google’s DeepMind learned how to play chess in just four hours before taking on another highly regarded chess computer, Stockfish8. In 100 games, AlphaZero won 28 and managed to lose none.
Eureka reports that while machines have obviously been showcasing their chess skills for a great number of years now, the common view in the past had been that their plays have been generally bland. However, researchers following AlphaZero noted how the computer showed creativity and even style in how it approached the games, making interesting moves that some argued humans would never consider.
The latest of many breakthroughs
This element of intelligence is being viewed as a watershed moment, which is particularly notable considering the huge breakthroughs seen in relation to the interaction of artificial intelligence and gaming down the years. Games such as chess have been adopted by AI researchers as excellent opportunities to train and observe AIs. They also serve as easy-to-understand and to explain indications of what said AI can do in the eyes of the general public.
The idea of using machines to dive into game theory has been around for a great number of years. In March of 2016, an AI developed by Google and called AlphaGo beat one of the world’s best players, Lee Se-dol in the ancient board game Go 4-1 in a tournament that made headlines around the world. The South Korean master of the game said that his defeat made him realise he needs to study the game more. Other well-known instances include Edward O Thorp’s work with an IBM 704, which played an important part in his understanding of the mechanics of blackjack in the 1960s and 1970s and helped him devise a detailed strategy for when to hit and when to stand as well as other considerations such as when to split pairs depending on card values. This strategy was elaborated in a best-seller book titled Beat the Dealer. Other developments saw the Canadian computer Chinook not only win the Checkers World Championship in 1994 but then go on to “solve” the game years later by confirming that if opponents play perfectly, a match will always end in a draw.
But what is it that makes a certain type of game so suited to play by an artificial intelligence? Well, one thing that the likes of chess, checkers and Go all have in common is that they all fit the criteria of a game of perfect information. Such a term harks back to the concept of game theory and the idea of using mathematical models to look at how participants – or players – in a set model of circumstances will behave and the decisions they will make.
In a game of perfect information, the players involved have access to all of the same details and strategies, and there is no information that is only accessible to one of the players. Therefore, they have a clear overview of how the action is unfolding in the game in question. This of course fits chess, as the two players involved can see the board and the location of each of the pieces, which means they can therefore make assumptions on the other competitor’s strategy based on what is happening in front of them. Quite simply, there is nothing hidden and there are no tricks up a player’s sleeve.
In comparison, a game of incomplete information takes place under a different set of circumstances and often with – unsurprisingly, considering the name – some element of information being held back from those involved. This may often include key detail which would help a competitor form an understanding of another player’s behaviours if you had access to it. An example of this is a card game where players can see their own cards but not the opponents’.
The latter description could fit a host of games that humans can easily participate in, but the circumstances of such gameplay tend to make it a lot more difficult for a computer or AI to get involved. The fact that nothing is hidden in chess and similar games makes it a perfect option for computers to play and attempt to master. However, that is not to say that efforts are not being made for AI to master types of games where not all of the information is available either.
In January of 2017, a competition was held in Pittsburgh which put four top poker pros up against an AI called Libratus, which was created by Carnegie Mellon University. A total of 120,000 hands were set to be held at the Heads-Up No-Limit Texas Hold’em event. The event was expected to be watched closely as, due to the nature of the game, the AI would only be able to be successful it had mastered a range of skills including one of the most human parts of this game – the bluff. Remarkably, the AI led the pros by a collective $1.7 million of chips after the last hands were played. The aim was not just to prove the AI’s poker prowess, however, as researchers are hopeful that developments in this area could be transferred elsewhere. For example, Carnegie Mellon University has suggested such bluffing powers could be used in an app to negotiate the price of goods or services.
Developments like the research involving Libratus and the groundbreaking work undertaken with AlphaZero are a clear sign that while Deep Blue in the 1990s was a landmark moment, work has continued in earnest to take artificial intelligence to the next level. But what is also clear is that the breakthroughs made in AI and games like chess have much wider implications, as it happened with our new-found understanding of checkers after AI solved it. It will be exciting to see what this fascinating relationship will continue to inspire.