I have noticed that during the server analysis gaviota (5 pieces) endgame tables are not used. They are available and their source of truth is shown during the client analysis.
Is there any reason the server analysis does not use it? It looks to me that there will be a win from all perspectives:
- more exact endgame evaluation.
- faster evaluation (gaviota probing takes ~0.1 s to give exact DTM, whereas server analysis uses 1 sec to give approximation)
- it is easy to implement. Check if 5 pieces, then use gaviota.
It would require all the clients that are providing the server analysis to have the full set installed, which is prohobitive. The server analysis is AFAIK provided by volunteers.
Moreover, Gaviota is now considered rather outdated, and Stockfish uses Syzygy format (much more compact and 50-move rule correct).
Syzygy is 6 pieces.
It would not be necessary to install on all clients.
If 6 pieces then direct to client with Syzygy.
This would free clients from analysing 6 men endings.
@Garff Gaviota and Syzygy are very different things. Syzygy (6 pieces) does not give you DTM, only DTZ and WDL.
As for full client install: it is not needed. As mentioned by
@tpr you can forward 5 pieces positions to clients which do have Gaviota. If you do not have such clients, it is possible to spawn one server which will be responsible for only 5 pieces endgames. This will be still more efficient: even on below average machine, python wrapper can probe Gaviota in ~0.1 sec.
@nikolajtesla The problem with DTM is that it gives wrong information for endgames affected by the 50 move rules. A DTM 60 endgame can be a draw or a win according to chess rules. That's why I'm saying Gaviota is outdated. Maybe it's a minor issue for 5 pieces only as there's less of such endgames in there.
Syzygy gives you DTZ50 which does tell you correctly whether a position is won or lost. That's why Stockfish only supports those. (And you can use 5 men just as well, it's much smaller than a Gaviota set). But it won't give immediate DTM announcements.
Fast or correct, choose one.
Author of Syzygy bases is interested in generating DTM syzygy at approximately 50% size of Nalimov bases.
@Garff it is not hard to combine it with Syzygy. Both tables are < 200Gb of data, which is laughable in terms on current hdds. Also lichess already stores them somewhere because the data is available though get requests.
All what is needed is to redirect all 5 or less endgame evaluations to that server.
IMHO it could be useful to install 5-man (or better if space allows) Syzygybases on all fishnet servers. However, configuration and verification (testing, if any is required) are nontrivial.
"All what is needed is to redirect all 5 or less endgame evaluations to that server."
But 5-man endgames often arise from positions containing more than 5 pieces! I'd argue that 99% of the Elo gain from having N-man Syzygybases installed comes from positions with more than N pieces.
Play and analysis are different.
In play Stockfish analyses until it reaches the 6 men Syzygybases and there comes most of the Elo gain from, agreed.
For analysis it is pointless to calculate a 6 men position as it is A) often wrong, B) consumes time. So if analysis of a 6 men position is requested it saves time and increases accuracy to redirect the request to a server that hosts the Syzygybases.
@Toadofsky I do not mean to use endgame tables to help to evaluate leaves of other positions. I mean to use endgame tables to only evaluate positions below 5-6 pieces.
For example in this game:
lichess.org/0Lgxuq0J/black#95 every position started from half-move 95 can be evaluated from the endgame tables.
The same for this game
lichess.org/UchpVBSV/black#131 starting from half-move 131
