What do you think?
Rate this book
Game Changer: AlphaZero's Groundbreaking Chess Strategies and the Promise of AI
It took AlphaZero only a few hours of self-learning to become the chess player that shocked the world.
The artificial intelligence system, created by DeepMind, had been fed nothing but the rules of the Royal Game when it beat the world’s strongest chess engine in a prolonged match. The selection of ten games published in December 2017 created a worldwide sensation: how was it possible to play in such a brilliant and risky style and not lose a single game against an opponent of superhuman strength?
For Game Changer, Matthew Sadler and Natasha Regan investigated more than two thousand previously unpublished games by AlphaZero. They also had unparalleled access to its team of developers and were offered a unique look ‘under the bonnet’ to grasp the depth and breadth of AlphaZero’s search. Sadler and Regan reveal its thinking process and tell the story of the human motivation and the techniques that created AlphaZero.
Game Changer also presents a collection of lucidly explained chess games of astonishing quality. Both professionals and club players will improve their game by studying AlphaZero’s stunning discoveries in every field that matters: opening preparation, piece mobility, initiative, attacking techniques, long-term sacrifices and much more.
The story of AlphaZero has a wider impact. Game Changer offers intriguing insights into the opportunities and horizons of Artificial Intelligence. Not just in solving games, but in providing solutions for a wide variety of challenges in society.
With a foreword by former World Chess Champion Garry Kasparov and an introduction by DeepMind CEO Demis Hassabis.
Matthew Sadler (1974) is a Grandmaster who twice won the British Championship and was awarded an individual Gold Medal at the 1996 Olympiad. He has authored several highly acclaimed books on chess and has been writing the famous ‘Sadler on Books’ column for New In Chess magazine for many years. Natasha Regan is a Women’s International Master from England who achieved a degree in mathematics from Cambridge University. Matthew Sadler and Natasha Regan won the English Chess Federation 2016 Book of the Award for their book Chess for Life.
The artificial intelligence system, created by DeepMind, had been fed nothing but the rules of the Royal Game when it beat the world’s strongest chess engine in a prolonged match. The selection of ten games published in December 2017 created a worldwide sensation: how was it possible to play in such a brilliant and risky style and not lose a single game against an opponent of superhuman strength?
For Game Changer, Matthew Sadler and Natasha Regan investigated more than two thousand previously unpublished games by AlphaZero. They also had unparalleled access to its team of developers and were offered a unique look ‘under the bonnet’ to grasp the depth and breadth of AlphaZero’s search. Sadler and Regan reveal its thinking process and tell the story of the human motivation and the techniques that created AlphaZero.
Game Changer also presents a collection of lucidly explained chess games of astonishing quality. Both professionals and club players will improve their game by studying AlphaZero’s stunning discoveries in every field that matters: opening preparation, piece mobility, initiative, attacking techniques, long-term sacrifices and much more.
The story of AlphaZero has a wider impact. Game Changer offers intriguing insights into the opportunities and horizons of Artificial Intelligence. Not just in solving games, but in providing solutions for a wide variety of challenges in society.
With a foreword by former World Chess Champion Garry Kasparov and an introduction by DeepMind CEO Demis Hassabis.
Matthew Sadler (1974) is a Grandmaster who twice won the British Championship and was awarded an individual Gold Medal at the 1996 Olympiad. He has authored several highly acclaimed books on chess and has been writing the famous ‘Sadler on Books’ column for New In Chess magazine for many years. Natasha Regan is a Women’s International Master from England who achieved a degree in mathematics from Cambridge University. Matthew Sadler and Natasha Regan won the English Chess Federation 2016 Book of the Award for their book Chess for Life.
416 pages, Paperback
First published January 25, 2019
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 30 of 34 reviews
December 18, 2021
[Original review, Mar 1 2019]
This book is something completely new in my experience, though I suspect we'll soon see more of them: it's a biography of an artifical intelligence. The entity in question is Deep Mind's AlphaZero chess engine, which was unveiled a little more than a year ago. For people who aren't familiar with the background, AlphaZero differs radically from all previous strong artificial chess players (there were a few more or less unsuccessful experiments, which will no doubt get appropriate footnotes in due course). Until late 2017, the only known way to create a powerful chess bot was combining efficient search of a huge number of positions with clever implementation of human chess expertise. Over the last sixty years, these methods had evolved engines, the best of which is Stockfish, that are far stronger than the top human players. IBM's Deep Blue famously beat World Champion Garry Kasparov in their 1997 match, winning the decisive last game when Kasparov gambled by playing an unsound opening. No one plays this kind of match any more: there's no point, since everyone knows the machines will win. The chess world had got used to the computer playing style. It looks inelegant and illogical, but experience seemed to show that it was unbeatable; chess, we all thought, did not really involve deep understanding, but mostly accurate calculation. Although Stockfish was known to have a few blind spots, it was regarded as being an almost perfect chess player in most normal positions, and the top grandmasters routinely evaluated their own play by checking it against the machine.
AlphaZero has at a stroke upended all that received wisdom. It doesn't search through a huge number of positions, and contains no handcrafted human expertise; instead, it's learned everything it knows by playing games against itself, using deep learning methods to train a neural net. It turns out that this is a much better approach. After just a few hours of training (admittedly on a cluster of several thousand processors), AlphaZero effortlessly trounced Stockfish. Over a thousand games, it won 155 and lost only six, most of them due to trying too hard to win in drawn positions. The most remarkable thing, however, was not the result, but its way of achieving it. AlphaZero's style is bold, imaginative, and, I'm afraid it's the appropriate word, beautiful. It is a pleasure to play through its games. They are extraordinarily aesthetic, and several of them have already become recognised classics. This book, written by a strong grandmaster who has spent months analysing AlphaZero's matches, gives you the details. It's very well done. I was impressed already, just on the strength of skimming through some of the games that are available on the web. Having now read GM Sadler's detailed analysis, my estimation of AlphaZero goes up even further. This is not just a strong chess engine. It is a non-human agent who, on its own and in less than a day, has discovered some extremely deep and interesting things about a game that people have been playing for over a thousand years. It's kind of incredible.
AlphaZero's most significant finding is that the initiative is far more important than we'd previously understood. It often sacrifices pawns and even pieces to get the initiative, without having calculated any kind of forcing sequence that directly justifies the investment. It can just see that the activity is more important than the material; if it were human, we'd call it intuition, but here that "intuition" comes from the trained neural net. In the same way, it has an extraordinary appreciation of how to manoeuver its pieces; again, if it were human, we'd call it "strategic planning ability". Sadler evidently thinks there's no point in adding twenty sets of scare quotes on every page, and just straightforwardly describes AlphaZero's thinking in these terms: it wants to put its knight here, it forces its opponent to move this pawn forward to free up the square it will use later on, it methodically eliminates Black's counterplay before it starts the decisive attack. As he says more than once, it's completely different from previous engines, who never teach you anything more interesting than that you should avoid making tactical errors. AlphaZero evidently has a very deep understanding of chess, and you can learn all sorts of important things from it. With my chess player hat on, I agree with him: this is the only sensible way to describe what's going on. With my computer scientist hat, I'm dying to know more. How on earth is this possible? I'd love to see further details.
In general, this book, fascinating as it is, just scratches the surface, and if you're also interested in both games and AI it will leave your head buzzing with urgent questions. Is the chess version's playing style an inevitable result of the training process, i.e. would it play the same way if you retrained it from scratch? Is the Go version's playing style fundamentally similar to the chess version's? (Alas, my Go skills aren't really good enough to tell). Could you extend these methods to play bridge, which is now the big challenge? Bridge involves uncertainty and communication; but I think I see ways to address those problems. Looking a bit further ahead, can you apply these methods to the real world, in particular to the problem of achieving goals that involve using language to interact with people? That's harder, but again you can see angles. If I had to make a guess, I'd say that we'll be there within ten years.
The Singularity is getting closer. Read Game Changer if you want to improve your view of what's coming.
____________________
[Update, Apr 8 2019]
You certainly get the impression that the top players are doing their best to learn from AlphaZero. Carlsen versus Giri yesterday was a particularly clear example. White against the current world #4, the World Champion played a quiet opening, then suddenly sacrificed a pawn to get a huge attack seemingly out of nowhere, using AlphaZero's trademark combination of an open file and an open diagonal both pointing towards Black's king. Giri's position fell apart and he also ran short of time; Carlsen could have played many fancy moves to try and win directly, but preferred the cold-blooded approach of taking the exchange and converting to a hugely advantageous ending. This proved effective, and Giri gave up in a hopeless position just before the time control.
____________________
[Update, May 15 2019]
AlphaZero's kid sister, LeelaChessZero, is playing Stockfish in the superfinal of the World Computer Chess Championship: after 24 of 100 games, she's 4-1 up with 19 draws. More details here.
____________________
[Update, May 23 2019]
After 71 games, Leela is up 10-5 with 56 draws and virtually certain to win the match.
It's interesting to watch on the TCEC page. You're shown both of the players' views of the position in real time, and you can contrast their respective takes on what's going on. Quite often, the assessments diverge significantly, with one machine thinking it has a large advantage while the other believes the position is about equal.
I have never found it meaningful to watch a computer-computer match before, but the AlphaZero revolution has shaken things up. The players have very different playing styles - suddenly there's an element of drama, and people care about what's going to happen. It's noticeable that there are a lot more Leela fans in the chat thread. And why not? She has played some excellent games, games that any human player would have been proud of. I was particularly impressed by #36.
____________________
[Update, May 26 2019]
Game #87 (in progress as I write) offers an interesting example of the two machines arriving at different conclusions about a position. Leela, Black, plays an offbeat opening where she exchanges her fianchettoed bishop early to spoil White's pawn-structure and block the centre. After 13 moves, Stockfish thinks he's better. He's got the two bishops, more space and better development, and figures that adds up to a 24% chance of winning. Leela disagrees. She thinks she's solid and that White has no long-term plan to improve his position, so she's the one with small but non-zero winning chances.
I'll post again when the game is finished.
____________________
[Update, later on May 26 2019]
To my surprise, Stockfish won game #87. It had no play, and Leela could just manoeuvre around to her heart's content and prepare a break. But she got the timing wrong, and when she did finally go for it the opening of the position turned out to be to her opponent's advantage. His bishops came alive and he soon had a crushing attack.
A remarkably human story in a fight between two machines!
____________________
[Update, May 28 2019]
The king is dead, long live the queen! Leela wraps up the match (final score: 14-7 with 79 draws), and is now the official Computer World Champion. In the TCEC chat thread, I see people gossiping about two more neural net players that have recently appeared.
____________________
[Update, Jul 2 2019]
Having just watched Alita: Battle Angel on the plane. I can't help thinking that Alita fights just like AlphaZero plays chess. No human being could achieve this level of power and precision, and it's very aesthetic to watch. The difference is that AlphaZero is real.
____________________
[Update, Oct 2 2019]
Game Changer has won the prestigious English Chess Federation Book of the Year award. The judges were unanimous. From their report:
[Update, Dec 18 2021]
An interesting passage from Sadler in the latest New In Chess Yearbook:
This book is something completely new in my experience, though I suspect we'll soon see more of them: it's a biography of an artifical intelligence. The entity in question is Deep Mind's AlphaZero chess engine, which was unveiled a little more than a year ago. For people who aren't familiar with the background, AlphaZero differs radically from all previous strong artificial chess players (there were a few more or less unsuccessful experiments, which will no doubt get appropriate footnotes in due course). Until late 2017, the only known way to create a powerful chess bot was combining efficient search of a huge number of positions with clever implementation of human chess expertise. Over the last sixty years, these methods had evolved engines, the best of which is Stockfish, that are far stronger than the top human players. IBM's Deep Blue famously beat World Champion Garry Kasparov in their 1997 match, winning the decisive last game when Kasparov gambled by playing an unsound opening. No one plays this kind of match any more: there's no point, since everyone knows the machines will win. The chess world had got used to the computer playing style. It looks inelegant and illogical, but experience seemed to show that it was unbeatable; chess, we all thought, did not really involve deep understanding, but mostly accurate calculation. Although Stockfish was known to have a few blind spots, it was regarded as being an almost perfect chess player in most normal positions, and the top grandmasters routinely evaluated their own play by checking it against the machine.
AlphaZero has at a stroke upended all that received wisdom. It doesn't search through a huge number of positions, and contains no handcrafted human expertise; instead, it's learned everything it knows by playing games against itself, using deep learning methods to train a neural net. It turns out that this is a much better approach. After just a few hours of training (admittedly on a cluster of several thousand processors), AlphaZero effortlessly trounced Stockfish. Over a thousand games, it won 155 and lost only six, most of them due to trying too hard to win in drawn positions. The most remarkable thing, however, was not the result, but its way of achieving it. AlphaZero's style is bold, imaginative, and, I'm afraid it's the appropriate word, beautiful. It is a pleasure to play through its games. They are extraordinarily aesthetic, and several of them have already become recognised classics. This book, written by a strong grandmaster who has spent months analysing AlphaZero's matches, gives you the details. It's very well done. I was impressed already, just on the strength of skimming through some of the games that are available on the web. Having now read GM Sadler's detailed analysis, my estimation of AlphaZero goes up even further. This is not just a strong chess engine. It is a non-human agent who, on its own and in less than a day, has discovered some extremely deep and interesting things about a game that people have been playing for over a thousand years. It's kind of incredible.
AlphaZero's most significant finding is that the initiative is far more important than we'd previously understood. It often sacrifices pawns and even pieces to get the initiative, without having calculated any kind of forcing sequence that directly justifies the investment. It can just see that the activity is more important than the material; if it were human, we'd call it intuition, but here that "intuition" comes from the trained neural net. In the same way, it has an extraordinary appreciation of how to manoeuver its pieces; again, if it were human, we'd call it "strategic planning ability". Sadler evidently thinks there's no point in adding twenty sets of scare quotes on every page, and just straightforwardly describes AlphaZero's thinking in these terms: it wants to put its knight here, it forces its opponent to move this pawn forward to free up the square it will use later on, it methodically eliminates Black's counterplay before it starts the decisive attack. As he says more than once, it's completely different from previous engines, who never teach you anything more interesting than that you should avoid making tactical errors. AlphaZero evidently has a very deep understanding of chess, and you can learn all sorts of important things from it. With my chess player hat on, I agree with him: this is the only sensible way to describe what's going on. With my computer scientist hat, I'm dying to know more. How on earth is this possible? I'd love to see further details.
In general, this book, fascinating as it is, just scratches the surface, and if you're also interested in both games and AI it will leave your head buzzing with urgent questions. Is the chess version's playing style an inevitable result of the training process, i.e. would it play the same way if you retrained it from scratch? Is the Go version's playing style fundamentally similar to the chess version's? (Alas, my Go skills aren't really good enough to tell). Could you extend these methods to play bridge, which is now the big challenge? Bridge involves uncertainty and communication; but I think I see ways to address those problems. Looking a bit further ahead, can you apply these methods to the real world, in particular to the problem of achieving goals that involve using language to interact with people? That's harder, but again you can see angles. If I had to make a guess, I'd say that we'll be there within ten years.
The Singularity is getting closer. Read Game Changer if you want to improve your view of what's coming.
____________________
[Update, Apr 8 2019]
You certainly get the impression that the top players are doing their best to learn from AlphaZero. Carlsen versus Giri yesterday was a particularly clear example. White against the current world #4, the World Champion played a quiet opening, then suddenly sacrificed a pawn to get a huge attack seemingly out of nowhere, using AlphaZero's trademark combination of an open file and an open diagonal both pointing towards Black's king. Giri's position fell apart and he also ran short of time; Carlsen could have played many fancy moves to try and win directly, but preferred the cold-blooded approach of taking the exchange and converting to a hugely advantageous ending. This proved effective, and Giri gave up in a hopeless position just before the time control.
____________________
[Update, May 15 2019]
AlphaZero's kid sister, LeelaChessZero, is playing Stockfish in the superfinal of the World Computer Chess Championship: after 24 of 100 games, she's 4-1 up with 19 draws. More details here.
____________________
[Update, May 23 2019]
After 71 games, Leela is up 10-5 with 56 draws and virtually certain to win the match.
It's interesting to watch on the TCEC page. You're shown both of the players' views of the position in real time, and you can contrast their respective takes on what's going on. Quite often, the assessments diverge significantly, with one machine thinking it has a large advantage while the other believes the position is about equal.
I have never found it meaningful to watch a computer-computer match before, but the AlphaZero revolution has shaken things up. The players have very different playing styles - suddenly there's an element of drama, and people care about what's going to happen. It's noticeable that there are a lot more Leela fans in the chat thread. And why not? She has played some excellent games, games that any human player would have been proud of. I was particularly impressed by #36.
____________________
[Update, May 26 2019]
Game #87 (in progress as I write) offers an interesting example of the two machines arriving at different conclusions about a position. Leela, Black, plays an offbeat opening where she exchanges her fianchettoed bishop early to spoil White's pawn-structure and block the centre. After 13 moves, Stockfish thinks he's better. He's got the two bishops, more space and better development, and figures that adds up to a 24% chance of winning. Leela disagrees. She thinks she's solid and that White has no long-term plan to improve his position, so she's the one with small but non-zero winning chances.
I'll post again when the game is finished.
____________________
[Update, later on May 26 2019]
To my surprise, Stockfish won game #87. It had no play, and Leela could just manoeuvre around to her heart's content and prepare a break. But she got the timing wrong, and when she did finally go for it the opening of the position turned out to be to her opponent's advantage. His bishops came alive and he soon had a crushing attack.
A remarkably human story in a fight between two machines!
____________________
[Update, May 28 2019]
The king is dead, long live the queen! Leela wraps up the match (final score: 14-7 with 79 draws), and is now the official Computer World Champion. In the TCEC chat thread, I see people gossiping about two more neural net players that have recently appeared.
____________________
[Update, Jul 2 2019]
Having just watched Alita: Battle Angel on the plane. I can't help thinking that Alita fights just like AlphaZero plays chess. No human being could achieve this level of power and precision, and it's very aesthetic to watch. The difference is that AlphaZero is real.
____________________
[Update, Oct 2 2019]
Game Changer has won the prestigious English Chess Federation Book of the Year award. The judges were unanimous. From their report:
The authors had the opportunity to understand the chess approach of AlphaZero and explore how it played in a series of games against Stockfish, one of the best human designed computer chess-playing programs. AZ won convincingly.____________________
Sadler looked at AZ’s games and found a unique style of play with many distinctive features, for example, piece activity, the initiative even at the cost of material and going after the enemy king, are just a few. Not unlike the young Tal perhaps, but more soundly based.
Chess players will find a splendid collection of games with many comparing AZ with players and games of the past. Very readable, there is nothing in the book that cannot be understood or enjoyed. The book is beautifully presented by New In Chess and is excellent value.
The ultimate accolade came from Carlsen who said that AZ had influenced his approach to chess. Game Changer may also influence yours.
[Update, Dec 18 2021]
An interesting passage from Sadler in the latest New In Chess Yearbook:
The set-up of the TCEC SuperFinal is quite unusual. The games are played at a long time control (120 minutes plus a 10-second increment) and all the games start from pre-determined opening positions. A match between Leela and Stockfish from the normal starting position would most likely end in 100 draws (yes, it has been tested!), so you need unusual and unbalanced openings to test both engines’ all-round capabilities and to provide entertainment (in the form of decisive results) to the watching chatters.
March 6, 2019
I ended up also getting the ForwardChess electronic version. It helped a lot since of much of this book is analysis of games. The e-version lets you play through on a board as you’re reading. As for the book itself, quite fascinating if you are interested in AlphaZero. Really amazing how it crushes Stockfish, sacrificing material left and right and locking up Stockfish’s pieces to the point where Stockfish just looks silly. And this is Stockfish we’re talking about!!
June 25, 2019
A great book. There is absolutely to many things too learn from the mighty Alpha-zero.
Although it took me so much time to finish this book but I think I have developed some great tactics in chess.
Although it took me so much time to finish this book but I think I have developed some great tactics in chess.
August 19, 2020
Chess is a personal favorite of mine. It doesn't matter to me that I am terrible at the game. If we were to go with ratings, then even at my best, I probably wasn't above a 2000 ELO rating. Computers are good at Chess. Any program could beat a Human World Champion. The set paradigm of Chess Engines focused on having a comprehensive opening book, brute-forcing the middle game, and having a giant book of Chess Endgames. The real situation is more complicated, but it involves decision trees and other mathematical ideas.
Artificial Intelligence and Chess have a history together. Human beings underestimate some things the brain can do and overestimate others. Take balancing, for example. A child can balance well enough to walk. A toddler is capable of seeing a picture and telling you what is in it. If you developed a computer that could do that, you would change the world.
Game Changer brings a new paradigm in Chess Engines. It talks about the Artificial Intelligence program called AlphaZero. To 'learn' the game, AlphaZero played against itself millions of times, gaining superhuman strength in 9 hours. The initial input only included the rules of the game. The book explores how the program thinks of the next move and applies a certain weight to each. This process leads to a lively attacking style that is very human-like.
The book has five main parts. Part one discusses the history of AlphaZero, and by extension, Computer Chess as a whole. Part two shows us what is happening when AlphaZero decides what move to make. This section is where mathematics comes in. Part three goes over the tactics of AlphaZero and how it compares to players on the Grandmaster level. It discusses the ideas it supports and how it comes to those situations. Part four reviews AlphaZero's opening move set. The final part of the book is the conclusion. It contains annotated games and interviews with the development team.
So whether you are a fan of Chess or have an interest in AI, this book has something to offer.
Artificial Intelligence and Chess have a history together. Human beings underestimate some things the brain can do and overestimate others. Take balancing, for example. A child can balance well enough to walk. A toddler is capable of seeing a picture and telling you what is in it. If you developed a computer that could do that, you would change the world.
Game Changer brings a new paradigm in Chess Engines. It talks about the Artificial Intelligence program called AlphaZero. To 'learn' the game, AlphaZero played against itself millions of times, gaining superhuman strength in 9 hours. The initial input only included the rules of the game. The book explores how the program thinks of the next move and applies a certain weight to each. This process leads to a lively attacking style that is very human-like.
The book has five main parts. Part one discusses the history of AlphaZero, and by extension, Computer Chess as a whole. Part two shows us what is happening when AlphaZero decides what move to make. This section is where mathematics comes in. Part three goes over the tactics of AlphaZero and how it compares to players on the Grandmaster level. It discusses the ideas it supports and how it comes to those situations. Part four reviews AlphaZero's opening move set. The final part of the book is the conclusion. It contains annotated games and interviews with the development team.
So whether you are a fan of Chess or have an interest in AI, this book has something to offer.
March 21, 2019
It's more of a chess book than an AI book. I like chess and Machine learning but I was looking for more AI.
August 28, 2019
Excellent book on how AI can approach playing chess. Instead of coding different weights, the AI discovers this rules for itself
March 24, 2019
I am a terrible chess player and pretty much have to take it on faith when moves are described as “stunning” and “amazing.” Still the excitement of the authors is evident and it’s hard not to be engaged by that.
May 27, 2019
Wait! Don't blow by this until you at least consider the ramifications. Yes, this is a geeky book for chess fanatics and I know that means most folks will not enjoy it. But...
Back in the mid 70's as I ambled through the Consumer Electronics Show I came across the first chess playing computer and played a game. My strength is about that of the average serious club player and I had no problem whipping it, finding it a bit clunky and slow. Still, I was impressed that it could handle a game that complex. Would a machine ever be able to beat a grand master? Not too likely.
Well we know how that turned out with IBM's Deep Blue beating world champion Garry Kasparov back in the 90's. Development has moved apace and these machines are playing with ratings level of about 3600 with the highest rated humans not even reaching 2900. These machines have their strength through their enormous computational power. Expert players and grand masters have provided instruction into how to evaluate moves and provided them with reams of "book" openings and tables of end game positions. The program then compares moves, evaluates them against each other in terms of how many pawns ahead or behind the resulting position shows and makes the choice. The ability to examine millions of options in a short period of time means a human will never ever again beat one of these programs over the board.
Well, along comes AlphaZero. What is different? The program was only provided the rules of chess and was left to run millions of games over 24 hours to teach itself the best way to play. No human input whatsoever with regard to strategy or tactics. What happened when it played the super programs like Stockfish or Houdini mentioned above? A huge number of draws naturally, some in games going over 256 moves, but in those games with a result AlphaZero crushed the opposition more than 80% of the time. The machine taught itself to a level of proficiency never seen before! (It also became the first machine to ever beat a "Go" master, an even tougher game than chess.)
Scary but awe inspiring. I read recently that a program was provided with photos of skin lesions and the associated diagnosis. When provided new samples, undiagnosed, the program was able to correctly identify the cancerous growths 75% of the time. Experienced and specialized physicians got it right 50% so the implications of Artificial Intelligence for other fields besides games are pretty impressive. Should be seeing some amazing things developed in the future!
Okay, you are free to go and ponder. The next remarks are for the chess players among you and probably hold little of interest for the rest. This book takes a long look at the AlphaZero games and analyzes them according to the various principles of play that enable it to win. Yeah, some it it looks pretty ugly to the avid chess player, but this machine seems to follow the idea of sacrifice, sacrifice whatever necessary to gain mobility for its pieces or restricting those of the opponent. It will sacrifice to do that as well as to open lines and control color complex squares. Many of the moves appear to have little purpose until the impact is revealed many moves later. The evaluative technique is based on win probabilities (like the Monte Carlo simulation used in financial analysis) rather than pure material like number of pawns up or down.
The two masters who authored the book provide a great human perspective and a lot of insight into the process. One thing that makes the book entertaining is that they will seek out historic games of famous masters and compare those games to ones the AlphaZero program developed.
For the serious chess enthusiast and computer geek, this is a "must own" book. I have not been as entertained by a chess work in a long time!
Back in the mid 70's as I ambled through the Consumer Electronics Show I came across the first chess playing computer and played a game. My strength is about that of the average serious club player and I had no problem whipping it, finding it a bit clunky and slow. Still, I was impressed that it could handle a game that complex. Would a machine ever be able to beat a grand master? Not too likely.
Well we know how that turned out with IBM's Deep Blue beating world champion Garry Kasparov back in the 90's. Development has moved apace and these machines are playing with ratings level of about 3600 with the highest rated humans not even reaching 2900. These machines have their strength through their enormous computational power. Expert players and grand masters have provided instruction into how to evaluate moves and provided them with reams of "book" openings and tables of end game positions. The program then compares moves, evaluates them against each other in terms of how many pawns ahead or behind the resulting position shows and makes the choice. The ability to examine millions of options in a short period of time means a human will never ever again beat one of these programs over the board.
Well, along comes AlphaZero. What is different? The program was only provided the rules of chess and was left to run millions of games over 24 hours to teach itself the best way to play. No human input whatsoever with regard to strategy or tactics. What happened when it played the super programs like Stockfish or Houdini mentioned above? A huge number of draws naturally, some in games going over 256 moves, but in those games with a result AlphaZero crushed the opposition more than 80% of the time. The machine taught itself to a level of proficiency never seen before! (It also became the first machine to ever beat a "Go" master, an even tougher game than chess.)
Scary but awe inspiring. I read recently that a program was provided with photos of skin lesions and the associated diagnosis. When provided new samples, undiagnosed, the program was able to correctly identify the cancerous growths 75% of the time. Experienced and specialized physicians got it right 50% so the implications of Artificial Intelligence for other fields besides games are pretty impressive. Should be seeing some amazing things developed in the future!
Okay, you are free to go and ponder. The next remarks are for the chess players among you and probably hold little of interest for the rest. This book takes a long look at the AlphaZero games and analyzes them according to the various principles of play that enable it to win. Yeah, some it it looks pretty ugly to the avid chess player, but this machine seems to follow the idea of sacrifice, sacrifice whatever necessary to gain mobility for its pieces or restricting those of the opponent. It will sacrifice to do that as well as to open lines and control color complex squares. Many of the moves appear to have little purpose until the impact is revealed many moves later. The evaluative technique is based on win probabilities (like the Monte Carlo simulation used in financial analysis) rather than pure material like number of pawns up or down.
The two masters who authored the book provide a great human perspective and a lot of insight into the process. One thing that makes the book entertaining is that they will seek out historic games of famous masters and compare those games to ones the AlphaZero program developed.
For the serious chess enthusiast and computer geek, this is a "must own" book. I have not been as entertained by a chess work in a long time!
March 23, 2022
I'm not strong enough at chess to appreciate everything in this book, but I definitely still got a lot out of it. (For reference, I'm currently rated about 1375 on Lichess, but I haven't been playing lately.) The authors do a great job pulling themes out of AlphaZero's games and finding interesting historical parallels to compare to.
I loved reading about positions that Stockfish and other engines thought were draws while AlphaZero could find a win. I also liked seeing the games with older engines in chapter 1, since they really highlight how far computers have come at chess.
My favorite part was learning about the algorithm AlphaZero uses to evaluate moves, and specifically the fact that when doing its Monte Carlo tree search, it doesn't use maximin for the lines it's analyzed. Instead, at each step, it takes a weighted average of the the possible responses for the opponent, weighting by its likelihood of that response being selected. At first, that sounded like playing "hope chess". If your opponents best response gives you an expected draw, but every other move would give you a strong advantage, then isn't it irrational to weight that move as better than draw based on the hope that the opponent will make a mistake? But I realized, this isn't purely hoping for a mistake from the opponent, but also realizing that your own algorithm could be over-valuing what you see as the best response. If you're sure about what the best response is, it's the only thing that matters, but if you're not, the fact that all the other lines are very good for you should increase your estimate of the position. In fact, it seems like this difference might be one of the main things allowing AlphaZero to "see danger" when the other engines see only draws. A human can look at a position and say "well, I don't see how to win yet for white, so it might be a draw, but I sure don't think black is going to win." AlphaZero can effectively make a similar evaluation, whereas Stockfish will evaluate as 0.00 until it actually finds a non-drawing line, which prevents it from distinguishing "safe" vs. "unsafe" moves in positions it thinks are drawn.
As the book went on, it got more into the weeds. My favorite was Part II, and honestly maybe I could have skipped Part IV. I was also a bit frustrated with how often the same game was discussed multiple times, and it wasn't even always called out clearly. I understand the authors wanted to organize the chapters by theme, and often one game had multiple themes. Even so, I would have preferred each game to be discussed only once.
I loved reading about positions that Stockfish and other engines thought were draws while AlphaZero could find a win. I also liked seeing the games with older engines in chapter 1, since they really highlight how far computers have come at chess.
My favorite part was learning about the algorithm AlphaZero uses to evaluate moves, and specifically the fact that when doing its Monte Carlo tree search, it doesn't use maximin for the lines it's analyzed. Instead, at each step, it takes a weighted average of the the possible responses for the opponent, weighting by its likelihood of that response being selected. At first, that sounded like playing "hope chess". If your opponents best response gives you an expected draw, but every other move would give you a strong advantage, then isn't it irrational to weight that move as better than draw based on the hope that the opponent will make a mistake? But I realized, this isn't purely hoping for a mistake from the opponent, but also realizing that your own algorithm could be over-valuing what you see as the best response. If you're sure about what the best response is, it's the only thing that matters, but if you're not, the fact that all the other lines are very good for you should increase your estimate of the position. In fact, it seems like this difference might be one of the main things allowing AlphaZero to "see danger" when the other engines see only draws. A human can look at a position and say "well, I don't see how to win yet for white, so it might be a draw, but I sure don't think black is going to win." AlphaZero can effectively make a similar evaluation, whereas Stockfish will evaluate as 0.00 until it actually finds a non-drawing line, which prevents it from distinguishing "safe" vs. "unsafe" moves in positions it thinks are drawn.
As the book went on, it got more into the weeds. My favorite was Part II, and honestly maybe I could have skipped Part IV. I was also a bit frustrated with how often the same game was discussed multiple times, and it wasn't even always called out clearly. I understand the authors wanted to organize the chapters by theme, and often one game had multiple themes. Even so, I would have preferred each game to be discussed only once.
May 27, 2019
i spent much of middle school really trying to learn about chess. i remember reading about the sicilian defense and pandolfini beginning chess. what I gathered out of that experience was that most of modern chess notation is utter crap and mostly unreadable. this book is a chess book and while it's ostensibly about a machine learning algorithm that learned something it's only about 10% information about that. as a chess book it's a very good one, showing where alphazero learned just how to take strategies that were extremely old at this point, rediscovering them on its own, then employing them in very aggressive ways. in that respect this book was enthralling. if you're willing to put up with the chess notation, you'll be rewarded.
March 24, 2019
A fascinating read. This book hits the intersection of several of my interests - chess, computer science, and machine learning. The potential of undirected machine learning is both exciting and a bit scary...
November 10, 2019
This is actually a remarkably good book, but my chess is nowhere near the level required to appreciate it fully. There are many pages of game analysis, some of which I replayed with a set at my elbow, and, Dear Reader, I ain't that good.
The chapters interviewing the architects, the historical anecdotes, and the technical commentary on AlphaZero, I found very rewarding - perhaps others less well tehnically versed might disagree ... not sure.
The vast superiority of the program over the preceding generation of human-thrashing software was well brought out: maybe there was a scope for a bit more about where such thinking will take us all?
Nice to see a picture of Judit - what became of Polgaria?
The chapters interviewing the architects, the historical anecdotes, and the technical commentary on AlphaZero, I found very rewarding - perhaps others less well tehnically versed might disagree ... not sure.
The vast superiority of the program over the preceding generation of human-thrashing software was well brought out: maybe there was a scope for a bit more about where such thinking will take us all?
Nice to see a picture of Judit - what became of Polgaria?
April 17, 2019
A lucid explanation of exactly what the difference is between AlphaZero's playing style and other chess engines like Stockfish and Houdini.
September 16, 2019
A close to perfect chess book. Fantastic organisation of themes. The right combination of analysis and exposition. Interesting sideline games between top players. Well written and engaging. It helps that the subject matter, being the rise of neural networks in AI (and in chess programming specifically) is both groundbreaking and, for me, fascinating.
May 22, 2020
Fascinating
Stockfish is one of the world’s leading chess engines, handcrafted over many years, well stocked with accumulated human learning on how to best play chess. It can easily beat all humans all the time including grandmasters and world champions. Even a resource depleted version running on a mobile phone is extremely formidable.
Alpha Zero is an AI / Machine Learning program which taught itself how to play chess without human input in nine hours.
Alpha Zero defeats Stockfish more or less easily most of the time. Moreover it plays in strange ways - which we can at least see the sense of after the fact but lack the means (circuitry) to make use of.
This ‘mind fragment’ is vastly superior to anything we will ever be able to accomplish in this domain. It will in due course be applied to other domains. Welcome to a very scary super high tech future, and not necessarily a benign one.
This book discusses the history of the development of ZeroAlpha, provides some insight into how it works and speculates as to the future. All very interesting.
BUT. It is primarily about chess. A huge part of the book is a discussion in detail about selected games between Stockfish and ZeroAlpha with extended commentaries from a human chess grandmaster. Set up a board. Play few at least a few of the games. It is a humbling experience. Watch on like a bug on the wall as two superhuman mind fragments battle it out. One even more superhuman than the other. Understanding some of it with help. But knowing they are playing at a level way beyond mere human comprehension and better than you will ever, even in your wildest dreams.
There is a great deal of discussion about Zero Alpha’s style of play and preferences. Savage attacks. Apparently wanton sacrifices. Mobility above much else. Gotta love it. But maybe something never to be directing more human activities.
If you are a chess enthusiast at any level this is a must read.
For everyone else a sobering look at the near future. They’re coming. They’re going to be much smarter than us in any way you care to define it. But will they be friendly? To which the obvious answer is, ‘why in the world would they be?’
Stockfish is one of the world’s leading chess engines, handcrafted over many years, well stocked with accumulated human learning on how to best play chess. It can easily beat all humans all the time including grandmasters and world champions. Even a resource depleted version running on a mobile phone is extremely formidable.
Alpha Zero is an AI / Machine Learning program which taught itself how to play chess without human input in nine hours.
Alpha Zero defeats Stockfish more or less easily most of the time. Moreover it plays in strange ways - which we can at least see the sense of after the fact but lack the means (circuitry) to make use of.
This ‘mind fragment’ is vastly superior to anything we will ever be able to accomplish in this domain. It will in due course be applied to other domains. Welcome to a very scary super high tech future, and not necessarily a benign one.
This book discusses the history of the development of ZeroAlpha, provides some insight into how it works and speculates as to the future. All very interesting.
BUT. It is primarily about chess. A huge part of the book is a discussion in detail about selected games between Stockfish and ZeroAlpha with extended commentaries from a human chess grandmaster. Set up a board. Play few at least a few of the games. It is a humbling experience. Watch on like a bug on the wall as two superhuman mind fragments battle it out. One even more superhuman than the other. Understanding some of it with help. But knowing they are playing at a level way beyond mere human comprehension and better than you will ever, even in your wildest dreams.
There is a great deal of discussion about Zero Alpha’s style of play and preferences. Savage attacks. Apparently wanton sacrifices. Mobility above much else. Gotta love it. But maybe something never to be directing more human activities.
If you are a chess enthusiast at any level this is a must read.
For everyone else a sobering look at the near future. They’re coming. They’re going to be much smarter than us in any way you care to define it. But will they be friendly? To which the obvious answer is, ‘why in the world would they be?’
November 15, 2020
An excellent book, complex but approachable. My brother (the chess player) sent it to me because I've done some work in AI. Although I couldn't follow the notations, the narration on how AlphaZero is changing the thinking of players is a fascinating read. I just happened to pick this book up before we started watching Queen's Gambit, which made it all the more enjoyable.
November 15, 2019
Outstanding. Instant classic for chess books. Illuminating expository writing and patient analysis. All studies of chess players should work like this instead of an endless string of chronological games!
July 7, 2019
Great book about the intersection of chess and AI.
Want to read
April 1, 2020Sample pages 19-21, 68-81, 196-201: https://www.newinchess.com/game-changer show some of AlphaZero's style.
March 16, 2024
AlphaZero is built by deep -mind , absolutely loved book for authors narration and demos of beautiful games ,.. AlpaZero is very Romantic in approach like Carl Von CLauswitz:)
4 principles that govern AI approach for AlphaZero (deep-mind)
- Learning over being programmed
- General rather than specific
- Grounded rather than logic driven
- Active rather than Passive
1. Learning rather than being programmed
The algorithm learns its strategy from examples rather than drawing on pre-specified human expert knowledge.
2. General rather than specific
The algorithm applies general principles and hence can be applied to multiple domains, e.g. shogi, Go, and chess.
3. Grounded rather than logic-based
Learning is based on concrete observations rather than preconceived logical rules.
4. Active rather than passive
The machine explores the game rather than being instructed by a human.
How Alpha zero training works :
Plausible move variation, selects from policy network based on probabilites (policy net)
evaluate variations and select better one , instead of taking principal variation (unlike other chess engines) it simulates and evaluates using MCST (Monte carlo search tree) and assigns value and learns into its value net
training of Alpha zero
Each time AlphaZero selects a variation to consider, it will be on the basis of three criteria:
1. how plausible the move is in this type of position (as determined by the policy network) ;
2. how promising is the outcome of the variation (as determined by the value network) ; and
3. how often this variation has been considered in the search.
other chess engines like stockfish position evaluation is linear based on some paratmeters unlike alphazero
https://hxim.github.io/Stockfish-Eval...
alpha zero considers many dynamic things(non linear) like restricted piece mobility which is difficult to interpret as neural networks form weights based on self trianing
AlphaZero's evaluation function is learnt and represented in terms of a neural network called the 'value network'
How does AlphaZero improve during training?
When it wins (or loses) a game, the connections in the value network are updated to evaluate each position in that game more positively (or negatively) . At the same time, the connections in the policy network are strengthened so as to play more often the move recommended by AlphaZero itself, after a lot of thinking by its Monte Carlo tree search. AlphaZero plays against itself millions of times, learning to provide better move suggestions (using the policy network) and to judge positions more reliably (using the value network) - essentially developing something akin to 'intuition' for how to play the game. This process of learning for itself, solely from its interactions, is known as 'reinforcement learning'.
It looks to us like AlphaZero uses piece mobility well and is a fantastic attacker. Do you think it looks at these concepts in a different way to Stockfish, perhaps more mathematically?
Those are well-known concepts in the computer chess literature, but in traditional chess engines they are usually applied with minimal or no selectivity. As a consequence, they have to be given low weights so that they do not exert an overly strong influence when their application is not justified. In the case of AlphaZero, the highly non-linear nature of neural networks means it can potentially learn to apply them much more selectively, and with higher influence where it thinks the features are valid. In addition, since AlphaZero maximises expected score, it is not so tied to keeping the material balance.
4 principles that govern AI approach for AlphaZero (deep-mind)
- Learning over being programmed
- General rather than specific
- Grounded rather than logic driven
- Active rather than Passive
1. Learning rather than being programmed
The algorithm learns its strategy from examples rather than drawing on pre-specified human expert knowledge.
2. General rather than specific
The algorithm applies general principles and hence can be applied to multiple domains, e.g. shogi, Go, and chess.
3. Grounded rather than logic-based
Learning is based on concrete observations rather than preconceived logical rules.
4. Active rather than passive
The machine explores the game rather than being instructed by a human.
How Alpha zero training works :
Plausible move variation, selects from policy network based on probabilites (policy net)
evaluate variations and select better one , instead of taking principal variation (unlike other chess engines) it simulates and evaluates using MCST (Monte carlo search tree) and assigns value and learns into its value net
training of Alpha zero
Each time AlphaZero selects a variation to consider, it will be on the basis of three criteria:
1. how plausible the move is in this type of position (as determined by the policy network) ;
2. how promising is the outcome of the variation (as determined by the value network) ; and
3. how often this variation has been considered in the search.
other chess engines like stockfish position evaluation is linear based on some paratmeters unlike alphazero
https://hxim.github.io/Stockfish-Eval...
alpha zero considers many dynamic things(non linear) like restricted piece mobility which is difficult to interpret as neural networks form weights based on self trianing
AlphaZero's evaluation function is learnt and represented in terms of a neural network called the 'value network'
How does AlphaZero improve during training?
When it wins (or loses) a game, the connections in the value network are updated to evaluate each position in that game more positively (or negatively) . At the same time, the connections in the policy network are strengthened so as to play more often the move recommended by AlphaZero itself, after a lot of thinking by its Monte Carlo tree search. AlphaZero plays against itself millions of times, learning to provide better move suggestions (using the policy network) and to judge positions more reliably (using the value network) - essentially developing something akin to 'intuition' for how to play the game. This process of learning for itself, solely from its interactions, is known as 'reinforcement learning'.
It looks to us like AlphaZero uses piece mobility well and is a fantastic attacker. Do you think it looks at these concepts in a different way to Stockfish, perhaps more mathematically?
Those are well-known concepts in the computer chess literature, but in traditional chess engines they are usually applied with minimal or no selectivity. As a consequence, they have to be given low weights so that they do not exert an overly strong influence when their application is not justified. In the case of AlphaZero, the highly non-linear nature of neural networks means it can potentially learn to apply them much more selectively, and with higher influence where it thinks the features are valid. In addition, since AlphaZero maximises expected score, it is not so tied to keeping the material balance.
January 23, 2020
As someone who really only plays chess every so often, I can’t say I got the most out of this book. I can however totally understand the appeal to much more enthusiastic chess players. For the most part, this book analyzes AlphaZero’s strategies and explains their brilliance. I’m guessing that much more experienced chess players can relate and understand AlphaZero’s flawlessness and impeccable execution much more than a casual player like myself. One thing that did stick with me was the fact that this is a completely self-taught AI. The thought of a singularity happening in which AI is above our own capabilities has lingered above my head for years, and it seems like it is now closer than ever. You can see things like AlphaZero, Deep Vision, and IBM Watson as sneak peeks of what could come in the future. Truly a world of possibilities lies beyond. This book captured me with its computer science and artificial intelligence, but could probably be magnified substantially with a proficient background in chess.
August 19, 2019
Analysis: No explanation of Stockfish or AlphaZero's flaws. Full game analysis is not given, just excerpts when AlphaZero is already winning, effectively being a book showing how AlphaZero converted already won positions, which tells us nothing. Then 5 pages of useless filler analysis showing how those games are won. This is surely Sadler's worst work, and I say this as a big fan of him.
Instructive Value: No new strategic or tactical ideas presented. Sadler did not know what he was teaching.
Coherence of Thesis: No working thesis. Nothing was changed, and this was just a long advertisement for Google, despite unfair match conditions to Stockfish. This is a very strange and unhelpful book.
Instructive Value: No new strategic or tactical ideas presented. Sadler did not know what he was teaching.
Coherence of Thesis: No working thesis. Nothing was changed, and this was just a long advertisement for Google, despite unfair match conditions to Stockfish. This is a very strange and unhelpful book.
February 19, 2022
In the days of its publishing, this book gave the chess players a chance to experience the same thing that AlphaGo created in the world of the go-game.
Considering that on the black & white side are two of the most horrifyingly strong silicon brains in the chess history, this books is arguably one of the most important game collection ever published.
However, the book is missing just a single small thing: the index of games. It is very hard to figure out how many games are covered, since some of the games are quoted more than once.
To solve that issue I have created a 2-page PDF (579kb) with the game-index (42 games!).
Whoever needs it, he/she could download it here:
https://www.dropbox.com/s/yi4zzexk50n...
Considering that on the black & white side are two of the most horrifyingly strong silicon brains in the chess history, this books is arguably one of the most important game collection ever published.
However, the book is missing just a single small thing: the index of games. It is very hard to figure out how many games are covered, since some of the games are quoted more than once.
To solve that issue I have created a 2-page PDF (579kb) with the game-index (42 games!).
Whoever needs it, he/she could download it here:
https://www.dropbox.com/s/yi4zzexk50n...
December 12, 2022
Not bad in theory, but I found the book too focused on chess and lost interest on the repetitive match analysis midway through.
I was looking for more applications, ways I can incorporate DeepMind and AI into games that matter to me, like Honmon, Pokémon VGC, Shogi/Go, or World Cup football. This book focuses too much on chess and certain chess celebrities. Magnus Carlson means nothing to me; I'd much rather hear from someone like me, a competitive player and programmer eager to incorporate AlphaZero techniques into my own AI code and gameplay strategy, not someone so far removed from intermediate layperson level.
I was looking for more applications, ways I can incorporate DeepMind and AI into games that matter to me, like Honmon, Pokémon VGC, Shogi/Go, or World Cup football. This book focuses too much on chess and certain chess celebrities. Magnus Carlson means nothing to me; I'd much rather hear from someone like me, a competitive player and programmer eager to incorporate AlphaZero techniques into my own AI code and gameplay strategy, not someone so far removed from intermediate layperson level.
August 24, 2019
This book is mixed bag. On one hand there are really some wonderful Alphazero games and concepts but on the other hand was not fan how material was divided and reading this book didn’t really give me an idea how Alphazero works. There could have been some simplified images presenting how it works and more games where familiar opening position was forced in order to see how Alphazeros handling of opening differs from other engines. Same with endgame handling.
May 18, 2020
Stealing from another reviewer, this is a biography of an AI entity, probably one of the first we will see in the future.
I dream of Chess books that don't require following board diagrams and I was hoping this would be more about ideas and stories than game analysis. It is not necessary to read it, but to fully grok the content, it is an assumption that you can follow Chess analysis through a few sparse board diagrams or have a board where you can play along.
I dream of Chess books that don't require following board diagrams and I was hoping this would be more about ideas and stories than game analysis. It is not necessary to read it, but to fully grok the content, it is an assumption that you can follow Chess analysis through a few sparse board diagrams or have a board where you can play along.
November 8, 2019
This book is almost completely chess analysis. I would have preferred a balance where the programming behind AlphaZero was also discussed at length. Some discussion of how AlphaZero was created and how it works is given but only in the most general of terms.
January 14, 2022
This is a great idea for a book and it tells an interesting story well, even if I think time has proven some of its conclusions to be a bit overstated.
Highly recommended for those with interest in chess and computers.
Highly recommended for those with interest in chess and computers.
June 9, 2019
Great book
April 12, 2020
a book how a chess new engine groundbreaking ideas in playing chess
Displaying 1 - 30 of 34 reviews