Game accuracy is highly dependent on how your opponent plays as well as ... of your opponent's playing strength relative to your own, ...
AlphaZero v Stockfish 2017
Sort:
Jan 7, 2023
0
#3
GM games are actually about 70% accuracy, even Bobby Fisher's average was 71% so it just really depends on your opponent.
Jan 7, 2023
0
#4
HendoTrees wrote:
GM games are actually about 70% accuracy, even Bobby Fisher's average was 71% so it just really depends on your opponent.
Where did you hear the 71% figure from?
Jan 7, 2023
0
#5
Taking some of the 1972 World Championship games and analyzing them at depth 30:
Game 3: 95.4% accuracy
https://www.chess.com/analysis/game/pgn/3pErXSVou8?tab=review
Game 4: 94.5% accuracy
https://www.chess.com/analysis/game/pgn/4bEoXf2itW?tab=review
Game 5: 97.9% accuracy
https://www.chess.com/analysis/game/pgn/4vQDC9gvZC?tab=review
Game 6: 91.6% accuracy
https://www.chess.com/analysis/game/pgn/3i3DRwmMWN?tab=review
All of these are in the 90s.
I've been reading "Game Changer" by Matthew Sadler and Natasha Regan - a book which describes AlphaZero's creation, how it plays, and its subsequent match against Stockfish.
One thing worth noting (that those who haven't read the book might not know) is that AlphaZero evaluates positions based on probabilities - not on accuracies.
It doesn't care if its moves are accurate or not; it only cares if its moves increase its probable chances of winning. And it learned that, to increase its chances of winning, it puts a large emphasis on piece mobility and kingside attacking chances.
As a result, AlphaZero is perfectly happy to play inaccurate moves, as long as they lead to open lines and open diagonals, pointing toward the enemy king.
This is different from how Stockfish evaluates a position - it tries to find the "best" move, based on centipawn calculations.
AlphaZero doesn't think in this way. It looks mainly at attacking chances via piece activity, and is happy to ignore any "correct" moves along the way, as long as it achieves its goals. So trying to look at the centipawn loss of its moves isn't really going to help much.
Jan 7, 2023
0
#7
Here's another one of their games analyzed to depth 30:
https://www.chess.com/analysis/game/pgn/4274jeNCt6?tab=review
93.8% v 90.5%
As with the other game, AlphaZero has a higher accuracy than Stockfish, but they are still not very high figures for machines that are above 3000 Elo
I have been watching Danny's lessons on AlphaZero and I took game 5 of AlphaZero v Stockfish 2017 and ran the game review analysis at depth 30.
https://www.chess.com/analysis/game/pgn/3U3KUnZPGA?tab=review
It gives accuracy percentages 88.9% for white and 87.7% for white, which is quite a bit lower than what I was expecting to see. Typically grandmaster games get percentages in the 90s and these engines are much higher rated than any human.
To what extent is the analysis wrong vs the engines at the time being bad compared to what engines know now? Is the depth that the engines were playing at a lot higher than 30?