Jan-04-19
 | | WannaBe: "Book", great annotation... |
|
Jan-04-19
 | | offramp: <WannaBe: "Book", great annotation...> It's like it's been annotated by a chicken.
Fritz 6 could be set to auto annotate games. Every threat of mate in one resulted in a compulsory annotation of "Can you see the mate threat?" by Fritz. Very annoying. |
|
Jan-04-19
 | | AylerKupp: This is what happens when chess engines don't have tablebase support. Although neither engine could have known it, the position is a theoretical tablebase draw after 71.Qxe4 per the Lomonosov tablebases. But after 79.Qxa5 the position is within Syzygy tablebase reach and the game is also a theoretical draw. Likewise after 93...Qxd5+ even though Black is a knight up. Instead the 2 engines, like drunken boxers, continued playing for another 64 moves. Maybe the engines were tired, given that this game is listed as round 1,224,984,858,386,794,290. But at least someone knew about the 50-move rule and called the game a draw after 143.Ne3. Otherwise the 2 engines might still be playing. |
|
| Jan-04-19 | | JimNorCal: I guess 43. ... QxN is avoided because White would have a perpetual? |
|
Jan-04-19
 | | keypusher: <AylerKupp> SF was using tablebases, according to the learned poster quoted below. Since A0 does not have a tablebase, perhaps Deep Mind wanted to see how it would do. AlphaZero (Computer) Or Deep Mind may just not have thought this through. There were quite a few oddities, including letting the engines play on in Q v. Q and R v. R endings. In one game A0 played a R+B v. R ending for dozens of moves (and of course got nowhere), then sacrificed the bishop (thus evading the 50 move rule) and played with R against R for 50 moves more. (It didn’t win.) My impression was that Deep Mind wanted to keep human intervention in the games to a minimum. They called draws for repetition, the 50-move rule, insufficient material (I think) and, as a final backstop, 512 half-moves — hence the numerous 256-move draws. They may just not know much about chess — A0 began as a Go program, remember. I think they didn’t start working with Sadler until after the games had been played. Separately on the A0 page I have a question for you about computing speed — would be glad to get your opinion if you get a chance. <WannaBe> The “book” annotation is useful because it idenfies games in which the engines used openings from the 2016 TCEC (engine) championships, as opposed to playing without an opening book at all. |
|
Jan-06-19
| | keypusher: <AylerKupp> Deep Mind were even worse about draws than I thought— they let this game go on 50 moves bishop v. bishop. Bishops of opposite colors, no less. AlphaZero vs Stockfish, 2018 |
|
Jan-06-19
| | AylerKupp: <<keypusher> SF was using tablebases, according to the learned poster quoted below.> Well, far be it for me to argue with such a learned poster. :-) Seriously, I keep getting the two matches mixed up; in the first match Stockfish did not have tablebase support but on the second match it did. Sorry about that. But it still doesn't explain why a game was not stopped once a position was reached that was a theoretical draw per the Lomonosov tablebases. I know that if I'm analyzing a game then as soon as the number of pieces remaining reaches 7 I consult the on-line Lomonosov tablebases to see what the theoretical result would be. Maybe the AlphaZero developers were hoping that Stockfish would make a mistake if the game continued for a while. But, with Syzygy tablebase support, once Stockfish found a 6-piece position during its search, it would not make a mistake. So if either program was going to make a mistake, it wouldn't have been Stockfish. So that possibility doesn't make sense to me. So I still don't understand why Deep Mind didn't adjudicate the position as a draw once 71.Qxe4 was played. |
|
|
|
|
Kibitzing
