There’s a new book out called Game Changer by Matthew Sadler, Natasha Regan about AlphaZero.
Google gave the authors unprecedented access and behind the scenes information about AlphaZero. The authors were given about 2000 games that AlphaZero played.
I’m generally not one to get excited about the insides of chess engines, but even I’m curious about AlphaZero and what makes it tick.
Read the book. The few games that were presented were at a fast time control. A limited Stockfish program vs. the AlphaZero program running on a powerful computer led to games that were shockingly or maybe not so shockingly bad depending on one’s perspective. It is a good propaganda book for Google and Deepmind, if you like propaganda.
Takeaways from the book. Pick your opponents. Make no mistakes. Play positional chess against the machines as evaluative functions are not as good as promoted. Some of the chess revelations from machine self-learning are not as impressive as it appears as one sees these ideas in top player games all of the time. Some examples are rook lifts and understanding positional compensation for a pawn sacrifice. Strong players know how much they can sacrifice, and like AlphaZero, do not work to get the material back as they are more concerned with speed of the attacking initiative, opening lines, and control of key squares or a color complex. This requires enormous effort of calculation and seeing deeply into the position, what is sometimes called intuition.
Computer programs do not “see” in the sense we do, but they are powerful and fast, calculating thousands of lines and can outmatch us in this regard. However, they are not as good in evaluating many types of positions and may place the correct candidate move lower down on its move selection list. Since humans have a variety of different thinking methods beyond tree analysis, through study and play experience we may have a deeper feel for position in less than ideal circumstances, that is, chaos on the board. This is hard for 99% of us to prove as consistency in move selection and play is hard to do and extremely tiring. We also attach emotional content to our moves which programs do not do. That can be both an advantage and disadvantage for machines, as when we discover something good the adrenaline and endorphin rush allows us to excel in pushing the envelope of our skills. For a machine a move is just a move. When Alekhine lifted a rook to the third rank, or sacrificed a pawn to clear a square, he was charged up emotionally by the process and knew he was correct without having to calculate deeply, even though he did it anyway. For a machine, there is no excitement that leads it to pick one move over another or immediately reject moves. Proving a machine mistake necessitates thinking and checking on its level which we have a hard time doing in a consistent manner. We think in chunks of concepts and ideas, not just individual moves.
Ultimately, the thinking processes developed by Deepmind for Google, will be used to sell us more stuff based on typical decision making patterns and specific data we have given away or been mined for. Chess, as usual, is just a thinking analogue for comparison. The processed data can be used to influence you as the corporations want to know what we do, how we think, what is attractive to us, and the types of stimuli that can be successfully targeted at us to make a decision that some other entity wants us to make. Oh sure, they will tell us they are using all of the data to cure cancer, eradicate poverty, and save the planet. The billionaires that own or will own all of the high tech in the future have all of our best interests at heart and not just massive profits and control of policy process to protect their interests, right?
Read the book. It is interesting, if choppily written. Don’t expect it to help your game very much.
Gustafsson’s chess24 Queen’s Indian repertoire (based on the cherry-picked first batch of games) is the most interesting attempt I’ve seen to draw general “human” conclusions from Alpha Zero’s play.
Being a tactical monster who already has a GM’s positional understanding helps… Carlsen’s recent play may have been partially inspired by Alpha Zero, but I agree that the book & the games are more of philosophical interest to OTB players below GM level.
Leela Zero is probably already a valuable tool for openings research and correspondence play.
Yes, some of the initial hype should be discounted, but I’d interject a minor quibble here (as a weakie relying on stronger players’ opinions). GMs seem to suspect that Alpha Zero was materially stronger even in the first batch of games. As you note, the Stockfish evaluation function is vulnerable, and deeper search in Stockfish isn’t much of a solution. It’s not as if fine-tuning Stockfish’s “value of a horse on e4” from (say) 3.24 to 3.34 will reduce the gap between the programs, it’s that Alpha Zero will generally evaluate the peculiarities of a position better than Stockfish. When there is not enough time to calculate a full solution, Monte Carlo’s returns will trump minimax after some point of diminishing returns. (There are exceptions: Stockfish was able to figure out how to break down the endgame fortress in Caruana-Carlsen London 2018(6). Alpha Zero couldn’t. But this is a position where minimax offers an absolute solution.)
I would like to be a lot stronger, but against similar or slightly higher *opponents, I’ve progressed to the point where I can sacrifice a pawn in a game for solid positional considerations. (Not gambit openings, but later in play).
I don’t have a percentage, because gambits don’t happen too often, but I won the last 2 games pretty solid. I have no problem sacrificing a pawn in a game, but I’m not going out of my way. I could go 2 days or 2 months between doing a positional sacrifice. -for purposes of this post, I’m only counting on positional gambits, where I’m doing something like freeing up lines of attack or just getting some more piece maneuverability.
Interesting takes on AlphaZero, you both had.
*perhaps up to 100 points higher than me, occasionally even a higher rated opponent.
What was the time control for the games between AlphaZero and Stockfish? That has a LOT to do with quality of games, even for computers. I think a game against Komodo, whichever the strongest version, against AlphaZero at classical time controls would better highlight AlphaZero’s playing ability. Too bad Shredder engine slipped so much. Perhaps a version 14 will come out sometime and be competitive again. Shredder, like Komodo, was more positional in it’s strengths.
The Wikipedia article for Monte Carlo tree search is a nice introduction.
Until reading Mr. Magar’s take on the book Game Changer I thought GM Jacob Aagaard’s review in the 2019/3 New In Chess magazine, still the best Chess magazine on the third rock from the sun, was the best review read by this reader. Basically, GM Aagaard rips the authors a new one in the nicest way possible. I purchased the book because of Aagaard’s review and am currently wading my way through it, while reflecting on how much more I enjoyed the FREE youtube videos by Sadler on the ‘games’ played by AZ and StockFish. I will reserve judgement until finished reading the book.
“A computer once beat me at chess, but it was no match for me at kick boxing.” - Emo Philips