This commit is contained in:
Miloslav Ciz 2024-12-30 20:45:53 +01:00
parent 9f0e34a0dd
commit 94fd1c5b4a
23 changed files with 1955 additions and 1939 deletions

View file

@ -316,11 +316,11 @@ What's the **theoretically worst game possible**, and how to find out? This is e
1/2-1/2
```
Lichess analysis seems to only handle the first 150 moves, the evaluation graph explodes up and down and almost jumps out of the roof. The following are the analysis results (for the first 150 moves). White: 15 inaccuracies, 15 mistakes, 97 blunders, 581 average centipawn loss, accuracy: 21%. Black: 11 inaccuracies, 17 mistakes, 97 blunders, 587 average centipawn loss, accuracy: 21%. That doesn't seem that bad, why aren't all moves blunders? Well, firstly the analysis is relatively quick (takes like 10 seconds for whole game), it likely doesn't see as deep as the engines who were given hours to play, but secondly we changed the rules of the game: the analyzing engine still assumes the players will be playing good moves, which is not the case.
Lichess analysis seems to only handle the first 150 moves, the evaluation graph explodes up and down and almost jumps through the roof. The following are the analysis results (for the first 150 moves). White: 15 inaccuracies, 15 mistakes, 97 blunders, 581 average centipawn loss, accuracy: 21%. Black: 11 inaccuracies, 17 mistakes, 97 blunders, 587 average centipawn loss, accuracy: 21%. That doesn't seem that bad, why aren't all moves blunders? Well, firstly the analysis is relatively quick (takes like 10 seconds for whole game), it likely doesn't see as deep as the engines who were given hours to play, but secondly we changed the rules of the game: the analyzing engine still assumes the players will be playing good moves, which is not the case.
For comparison here is another bad game in which we just take regular stockfish 17 and make moves like this: from all possible moves, minus the ones that draw, choose the one that leads to the position with worst evaluation for us. 3 seconds are given for evaluating each possible move, so we get something around a minute to make a move. For "mate in N" we take the move that gets us mated sooner as better, and to decide between several "mate in N" moves with same N we try to estimate the worst by taking an average static evaluation of the board to depth 3 (for technical reasons we use [smallchesslib](smallchesslib.md)'s evaluation) -- this should help us prefer positions in which there are more ways to get ourselves mated or in which we at least lost most material and other advantage on average. This game embodies the effort to make the worst blunder in each move in a regular game of chess -- as such we won't see too many "forced blunders", just great many generous offers that keep being turned down. In result this produced another terribly long game:
{ My computer basically spent the whole day computing this game instead of mining Monero, so please enjoy :D ~drummyfish }
{ My computer basically spent the whole day computing this game instead of mining Monero, so please enjoy :D NOTE: I don't actually mine Monero of course, I'm not stupid enough for that. ~drummyfish }
```
1. g4 f5 2. f3 g5 3. Kf2 Kf7 4. Ke3 Ke6 5. Kd4 Qe8 6. b4 Qh5 7. f4 Kf6 8. Ke3 Qh3+ 9. Kd4 Qc3+