All too human: Game Changer
reviewed by Ben Graff
A self-taught computer plays more like a human than we could ever have imagined. AlphaZero has crushed its conventional rivals and opened new frontiers for chess, as Matthew Sadler and Natasha Regan explore in this seminal book.
Human fascination with machines that can seemingly play chess dates to the eighteenth and nineteenth century. The Turk and later AJEEB appeared to be autonomous chess playing entities. Impressive though the engineering was, the skill lay in the illusion, both had a chess master hidden inside. The great Harry Pillsbury forced his six-foot frame into AJEEB’s cramped and musty confines for over a decade, apparently drinking a quart of whisky a day to relieve the pain. Perhaps a contributory factor to his death at the age of thirty-three. After the second world war paper-based programs played chess for the first time and by the 1970s chess computers were fully with us, even if they were all terrible.
It took a while to find a way to program computers such that they were not wholly materialistic. It is much easier for a computer to understand that a pawn is nominally worth a point than to place a realistic value on the control of a square or an open file. Especially when no fool proof formula exists to measure such apparently abstract gains. As computer processing power grew, the ability to hard code rules and data and the capacity for computers to calculate every possibility inevitably led to a huge increase in their strength.
In 1982, Grandmaster Jan Hein Donner was fortunate to stagger past Belle, the leading computer of its era. Donner did not play as well as he could have done, but in truth the writing was already on the wall. In 1997, world champion Garry Kasparov would succumb to the IBM machine Deep Blue, which may or may not have had a little help from its operators. Either way the era of debate as to which was stronger, man or machine, was drawing to a close. Humans could at least console themselves that these brilliant computers were still dumb. They followed the mass of rules they were programmed with and relied on their processing power to crunch every possible option. They had no instinctive understanding as to what was best. The computers might have been better than us, but they certainly were not intelligent and heck, their games were often ugly and difficult to learn from.
In contrast AlphaZero is a quite different animal, which is what makes Matthew and Natasha’s book so timely. Relying on artificial intelligence rather than brute force processing, Alpha Zero was given no more than the rules of chess and proceeded to play 44 million games against itself over the course of nine hours, in which time its playing standard progressed from hapless beginner to that of the most formidable chess playing entity on the planet.
There is something profoundly moving about a machine that starts out from nowhere, quickly surpasses all that has gone before, yet still manages to reaffirm much of what we know to be good along the way. I enjoyed the way in which Matthew and Natasha drew parallels between AlphaZero’s games and those of great players from the past and present including Chigorin (disregarding material imbalances) Alekhine (opening a second front in an endgame) and Polgar (using a queenside file to defend the kingside.) AlphaZero’s style appears to be much more human than the brute force processors it has surpassed. Its intuitively logical, attacking approach may be to a higher standard than that of any living and breathing (or other silicon) entity, but it is certainly relatable.
With a foreword by Garry Kasparov and an introduction from Demis Hassabis (AlphaZero’s creator) Game Changer is structured such that we learn about the history of the program, how it thinks, the key themes in its play (and fascinatingly what we can learn from them) and its openings. You sense Matthew and Natasha’s pioneering excitement, as they explore AlphaZero’s domination of Stockfish (the strongest conventional engine) in the games which make up the bulk of this book.
As Matthew puts it on first viewing: “As a chess fan, I was hoping to discover something original and fantastic; as a strong player I was prepared to be modestly disappointed by the style of play….I whizzed through the first ten games…but nothing yet to make me stop and gasp. The 11th and 12th games did. As the 11th game flashed before my eyes, it suddenly registered that AlphaZero as Black had already sacrificed three pawns and had just played the quiet move 22… h6 on the kingside.”
Matthew and Natasha do a great job of distilling AlphaZero’s approach into key principles from which we can learn. Some seem relatively obvious, such as king safety, finding strong outposts for knights (albeit incredible examples are given as to how AlphaZero creates these in practice) and ensuring that the centre is stable before attacking on the wing. Others, such as understanding when to make sacrifices for long term gains, attacking with opposite coloured bishops, delaying placing rooks on open files if other files around the opponent’s king might be openable, and utilising rook pawns perhaps less so. I was also stuck that AlphaZero does not appear to be a proponent for the minority attack.
AlphaZero’s intelligent, principles-based approach enables it to understand positions better than its data-crunching rivals. Matthew and Natasha highlight numerous examples where other machines will evaluate a position at 0.00 but AlphaZero will believe (and subsequently demonstrate) that it is ahead, often taking forward our understanding of the game with profound originality and imagination. As Natasha highlights, while we can learn a lot from AlphaZero, having the skill to play like it in practice is something else entirely!
There is much of interest in this book about the science behind AlphaZero. Anyone with a more general interest in Artificial Intelligence and its broader applications will doubtless get a lot from Game Changer. The chess games are fantastic, and Matthew and Natasha do a great job of both exploring and explaining Alpha Zero’s play in a way that players at all levels will be able to learn from.
If I had any minor quibbles, they really stem from the fact that this is a big, ambitious book with a lot of material in it. Part story, part AI primer, part terrific chess, part coaching manual; there is a lot to go at and occasionally the structure feels a little textbook like, but equally I am not sure how else the material could have been arranged. Readers should not be intimidated, but this is a book to savour over time rather than to be gulped down in a couple of sittings. It is certainly a rewarding experience.
Perhaps our move from The Turk and AJEEB, via brute force machines, to AlphaZero has seen humanity come full circle. Harry Pillsbury might not be hiding inside, but somehow his spirit and that of all chess players inhabits AlphaZero, with its capacity to teach us things we never imagined, in a style that makes it feel like one of our own.
Game Changer by Matthew Sadler & Natasha Regan – New In Chess 2019