An artificial intelligence can identify entertaining chess moves rather than just winning ones. This could lead to AIs that are more fun to watch or play against, and even offer insights into how we appreciate the game.
AI has long been able to beat even the most skilful human players at chess, with IBM’s Deep Blue famously defeating world champion Garry Kasparov almost 30 years ago. But AI tends to analyse the game and make the move that is most likely to win, with no consideration for entertaining spectators or executing an aesthetically pleasing tactic.
Now, Kamron Zaidi and at the University of Toronto in Canada have developed a test to identify “brilliant” moves. This is a common term in chess when an unexpected, creative or bold move that ultimately proves advantageous is played, and is marked in records with a “!!” notation.
Advertisement
Guerzhoy and Zaidi used data on 8574 games from the online chess website , which had 820 moves marked as brilliant, to train a neural network to decide whether a given move would be seen as brilliant by humans. When tested on real moves and real human reactions, it achieved this with 79 per cent accuracy.
The key to identifying brilliance lies in what is called a game tree, says Guerzhoy, which is an expanding web of possible moves that spread out from a given point in a match. Some game trees will include moves that are considered brilliant, while others will not.
The researchers found that a long, narrow path through a tree correlates with a move that is seen as brilliant. These might not be spotted by a typical AI chess model, which is likely to be analysing just a few moves ahead, and the long paths can look far further into the future and require more planning.
The pair also found that a tree containing a brilliant move may actually look much less than brilliant – or even terrible – until the final move. That surprise success may be a key part of its brilliance.
“If it’s the case that the path through the tree is really narrow, then that means you needed to have had a plan. The plan needed you to look really far ahead,” says Guerzhoy. “You’ll have to go really deep in the tree to see that this is actually a successful strategy. That is something that people would find to be kind of aesthetically appealing.”
Guerzhoy says the main aim of the research was to gain insight into how people perceive the aesthetics of chess, but it also paves the way for a chess AI that aims to achieve brilliant and impressive moves, rather than just win. The researchers hope to release just such a model publicly in the coming months.
Chess grandmaster says AI has had a profound impact on chess, with models running on even modest hardware now able to play with “superhuman” ability. This has changed the game, making some previously well-regarded tactics obsolete and throwing up many new ones. “AI’s taken away some stuff, but it’s also put a lot back in. An awful lot of stuff that we never even considered was possible has been revealed by [AI] engines,” he says.
Sadler believes that a model that can understand and emulate brilliance could be an essential training tool for professionals, and a more enjoyable opponent for amateurs.
“It kind of shows the right spirit, right? Because they’re trying to make AI understand how humans view things,” says Sadler. “I think that’s a small step along the way to making AI understand the context in which we play chess, which is totally different to the context in which AI plays chess.”
Guerzhoy says a similar approach may also shed light on what humans find brilliant in other areas, such as poetry or classical music, where there is a formal framework.
arXiv
