mmt Vs. Ovyron (G4 D5 BG2)

[zullil]

https://www.chessdb.cn/queryc_en/

You can explore the chess lines that it has and request analysis for the moves that it doesn't; it will expand the moves you insert automatically and backsolve their scores to propagate them backwards so previous positions have them ranked correctly in real time.

Thanks for the link. How good its results will be depends a lot on how deep its analysis is at leaf nodes, right (and that's unspecified)?

If this is what I think it is, I believe evals are from Stockfish searches, but at shallow depths (like 22?).

[jp]

If this is what I think it is, I believe evals are from Stockfish searches, but at shallow depths (like 22?).

Here's what noobpwnftw said about it previously.

There are two queues: one is for move exploration, each position is evaluated by a shallow SF search on multi-PV(depth 12), sieving from all moves for a minimum of 5 and within 200cp of the best score, these moves are then sent to the second queue; the second queue evaluates each position after making those moves by performing a normal depth 22 search, then without clearing the hash, searches the original position again and store the results to the database. Some back propagation work is done both during queries(look up to 20 plies deep), upon analysis request(100 plies deep) or in background.

Currently there are about 200 cores on the first queue, 4500 cores on the second and a few machines generating self-play games with various engines. Parameters are chosen from experience, engine code is at https://github.com/noobpwnftw/Stockfish/tree/siever. Due to the nature of such brute-forcing method, it does not require up-to-date everything to make it work, yet using the latest engine code can be more efficient.

I'm not sure what he meant by "from all moves for a minimum of 5".

[jdart]

chessdb.cn is quite a nice system, and has a remote API, so you can actually call this from an engine or GUI and use it as a book.

22-depth searches with backpropagation is a decent start at a book but it is not going to be perfect. GMs are certainly analyzing deeper than that and there are cases where the best move is something known from correspondence games where they have days to make a move.

[Ovyron]

Accepted.

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3

[d]3qkb1r/p2n1ppp/1rp1b1n1/3p4/3P4/2NQ1N1P/PP2PPB1/R1B1K2R b KQk -

(probably playing 12...Bd6)

[Ovyron]

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6

[d]3qk2r/p2n1ppp/1rpbb1n1/3p4/3P4/2NQ1N1P/PP2PPB1/R1B1K2R w KQk -

[mmt]

Now the game is becoming very interesting. Move 11 was one important choice for white. h3 is the move that seems to be preferred, but the difference with some other moves, like Nf3, isn't all that big, and it may turn out to be worse.

SF 11 mostly goes for h3 but it has Nf3 and even b3 as the top choices for a few plies. But LC0 sticks to h3 by around 0.20 pawns.

[jp]

3qk2r/p2n1ppp/1rpbb1n1/3p4/3P4/2NQ1N1P/PP2PPB1/R1B1K2R w KQk -

Komodo 13.2, depth 29: -0.97 h4.

(I meant to see depth 28, but overshot.)

[Alayan]

Ethereal 11.88 to depth 34 prefers h4 too, with a suspiciously low eval of -0.81. It would never hold this position against Stockfish.

[zullil]

3qk2r/p2n1ppp/1rpbb1n1/3p4/3P4/2NQ1N1P/PP2PPB1/R1B1K2R w KQk -

Komodo 13.2, depth 29: -0.97 h4.

(I meant to see depth 28, but overshot.)

At depth 44, Komodo has -1.09. I'll refrain from giving its current favorite for White's move 13, since White hasn't played that move yet.

[mmt]

A little update: I'm changing LC0's weights to net 59866. I haven't seen or done enough testing to know that my previous bigger Sergio's net is better. I'm also updating the parameters for wider search because I'm doing very long calculations and I'm using SF to guide LC0 if there are tactical problems. I will be recompiling LC0 to use doubles instead of floats like koedem has suggested - this seems more important in long searches.

h4 is likely next.