@zanycolourfulwindow I'm not expressing distaste with engines, I'm simply stating that the authors were making an assumption that given a set number of pieces (8) with at least 1 set of opposing pawns, they could always forecast if the game would be a win, lose, or draw. I am simply stating this statement is inaccurate UNLESS we are presuming the play with follows is perfect play. We cannot know the outcome of a game with these variables, UNLESS we are to presume that every move that follows will be the best one.
Should have sent the drives to me, I would have done it in a week :')
In all seriousness though, insane achievement!
Impressive stuff, and a good idea to use the op1 metric to make this more tractable. Can you expand a bit on why you think DTC is a better metric than DTZ50, to the point where you'd even like to redo the 7-man under DTC? I mean, my favourite would definitely be DTM50, but that just makes the search space explode, so DTZ50 seems to me to be the only realistic way of honoring the 50-move rule.
please make 10000000000000000000000000000000000000000000000000000000000000000000000000000000000 pc tb
if you wanted to do what @OjaiJoao said, you will have a hard time understanding.
@Sesse said:
> Impressive stuff, and a good idea to use the op1 metric to make this more tractable. Can you expand a bit on why you think DTC is a better metric than DTZ50, to the point where you'd even like to redo the 7-man under DTC? I mean, my favourite would definitely be DTM50, but that just makes the search space explode, so DTZ50 seems to me to be the only realistic way of honoring the 50-move rule.
For the purpose of perfect play, let's say for determining the truth-theoretical value of a position or just to finish off an engine or engine-assisted game, DTZ50 is sufficient.
However, the moves recommended by DTZ tables can often be weird because if any pawn push is available that keeps a position winning, it will be the table's top move. Likewise, the defensive moves in losing positions tend to be extra weird.
So from the perspective of tablebase as a teaching and analysis tool for human players, DTZ is less useful than DTC. I would guess that's the reasoning for the parts of the blog post you are referring to.
@IndigoEngun said:
> However, the moves recommended by DTZ tables can often be weird because if any pawn push is available that keeps a position winning, it will be the table's top move. Likewise, the defensive moves in losing positions tend to be extra weird.
Well, DTC is no less weird. If it can force the opponent to accept a piece sacrifice (into a still-winning ending, no matter how complicated), it will. In this aspect, it's exactly the same as DTZ. (DTZ is, of course, happier to push pawns before doing almost anything else, but it doesn't seem like much of an additional problem?) Only really DTM will do if you want “non-weird” play, and even then, using tablebases in search will give you weird effects as the engine is desperate to transition into tablebase land.
Yes, DTC has its own issues too, but that was my best guess has to why it's preferred over DTZ, which has DTC's issues and then more.
DTM would be better but it also increases the size of the tablebase several times over.
@Sesse said:
> Only really DTM will do if you want “non-weird” play, and even then, using tablebases in search will give you weird effects as the engine is desperate to transition into tablebase land.
That's completely true when you are analyzing positions with a few more pieces with the help of an engine equipped with TBs, but not so much when you directly look at the 8-men position itself, which I think is the main intended use case here on lichess.
> Can you expand a bit on why you think DTC is a better metric than DTZ50, to the point where you'd even like to redo the 7-man under DTC?
Marc had already computed all the tables as DTC before Lichess came into the picture. We just didn't want to risk too many hard drives at the same time, and skipped 7-piece tables for the first of potentially multiple transfers.
The result is that, for the time being, it's a bit jarring to analyse an endgame on Lichess and switch from DTC to DTZ50 half-way.
Personally, I slightly prefer DTZ50 with 5-valued WDL, because it's able to answer questions about "pure" chess, but *optionally* also the 50-move rule, all while the tables tend to compress better.
I'll have to ask if there's a deeper reason why op1 is DTC, or if it's just historical happenstance.
@EVmaximiser said:
> Another example, from the op1 set: 6k1/6b1/5p2/4nP2/8/8/4P3/3QK3 w - - 0 1: The position is a fortress, but Stockfish evaluates it as 1.51 after (2-)60 minutes of analysis (31 GN, depth 100). Leela gets this one right enough quickly, and goes further and further down in evaluation unlike Stockfish.
Similar position from a real game:
https://lichess.org/78p8YLIY#88
To be fair, Stockfish finds the better move 45.Kf3 +- and also other better moves earlier, but after 45.g4? = it still believes in an advantage of approximately 1.5 pawns.
@OjaiJoao said:
> If you can find positions from real games that can fool my complete centaur setup of strong human + SF + tablebases, I'll be really impressed. I doubt there are more than a tiny handful of them.
Add a good graphics card and Lc0 to that setup and you have a good chance of detecting an additional promille of the billions of fortresses in chess which no one will ever fully understand. Strictly spoken, every gambit where the win for the side with the material is not proven is a fortress. That is not a tiny handful.
https://lichess.org/analysis/8/kp2q3/1Np5/pPPp1b2/P2Pp1p1/4Pp1p/5PPB/1R4K1_w_-_-_0_1