Weiss 0.9 64-bit Gauntlet for CCRL 40/15

[Gabor Szots]

What size TBs did you use, 5- or 6-men?

6.

[Terje]

What size TBs did you use, 5- or 6-men?

6.

I see. I looked through about a third of the games, and all the time outs happened in 5/6-men positions. I have done practically no testing with Weiss having TBs but no TB adjudication, so I assume Weiss is just not handling searching while in a TB position well. I will look into fixing this.

[Terje]

Hi Terje,

I could not find a contact to you that's why I write here.

I have lots of time forfeits with 0.9. Weiss loses on time in approximately every 8 games. There was nothing wrong with 0.8. I read you'd modified time management but something must have gone wrong. I don't know if Graham encountered the same problem.
This is 2 minutes + 1 second increment time control. Unfortunately no logs.

Regards,
Gabor

In addition to the new time management in 0.10, I have now implemented TB probing in root so Weiss will get the move from the tablebase and play it right away without searching if the base position is in tablebase. On my machine this takes around 1ms so with the next release there should definitely not be any time issues in endgames. It should also no longer blunder (in tablebase positions)

[Gabor Szots]

In addition to the new time management in 0.10, I have now implemented TB probing in root so Weiss will get the move from the tablebase and play it right away without searching if the base position is in tablebase.

I'm not sure you can avoid searching in those cases. Playing the TB move immediately may not get you closer to mate and the 50-move rule may prevent winning.

[Terje]

In addition to the new time management in 0.10, I have now implemented TB probing in root so Weiss will get the move from the tablebase and play it right away without searching if the base position is in tablebase.

I'm not sure you can avoid searching in those cases. Playing the TB move immediately may not get you closer to mate and the 50-move rule may prevent winning.

Weiss plays the move that is fastest to zeroing the 50mr while still winning, or the slowest while drawing/losing. This results in some funny play, sacrificing material and/or taking enemy pieces that aren't relevant (because doing so zeroes the 50mr), but it keeps the correct result, even against Stockfish which does keep searching (in my limited testing).

https://syzygy-tables.info/?fen=8/2q5/6 ... _w_-_-_0_1
Giving SF white and Weiss black here results in getting mated in 50 without the pawn ever moving, as that resets the 50mr slower than letting white push the pawn. Swapping colors, SF gets mated in 86 after the pawn promotes and the two rooks are traded for the black queen. Clearly SFs defense is harder to beat for a human/engine without tablebases, but as long as both sides have them it makes no difference.