Terje wrote: ↑Sat May 30, 2020 1:59 pm
Given that a lot of engines have exactly 41 errors it would be interesting to know which positions (or even 1 of them) these occur in.
4Rk2/2R5/2p3p1/3p4/3P3r/P7/1P3P1p/7K b - - bm Kxe8; ce -32000; acd 0;
k5rb/3r4/1ppP4/p1p1p1p1/P1P1Pp1p/1P1R4/3R1PPN/3q1K2 w - - bm Rxd1; ce -32000; acd 0;
8/3P2k1/5p2/6pp/4p3/1b2P2P/6r1/2BK4 w - - bm Ke1; ce -32000; acd 0;
3r2k1/6p1/4p3/1p1nB1Qp/1PpP3P/R1P2P2/1r4qK/8 w - - bm Qxg2; ce -32000; acd 0;
'What they have in common, it's called the chopper move.
90% of coding is debugging, the other 10% is writing bugs.
Rebel wrote: ↑Sat May 30, 2020 3:29 pm'What they have in common, it's called the chopper move.
How can an engine have trouble selecting the one and only legal move in a position?
I'm guessing they select it and play it without giving an info string. Also not spending their time thinking, wasting that turns time which they could use to fill TT. Playing only-moves instantly is an error in 'go infinite' and a bad choice in 'go movetime'.
Confirmed to be the case in rofChade (works correctly in 'go infinite').
Comparing to the CEGT blitz list (I picked CEGT because it uses ORDO, which agrees with normal elo calculations for match results), the difference between Stockfish 11 and Fruite 2.1 (I picked it as it's near the bottom of your list, easy to find on cegt, and a very well known reference point) is 990 elo on CEGT, 532 elo on your list. So doubling your ratings and subtracting a constant would not be way off, but multiplying by 1.86 with a suitable subtraction would be ideal. Of course a proper analysis of all engines on both lists will give a somewhat different ratio, but probably not too far from this 1.86 value.
On all 3 rating lists the longer time control the more the elo of an engine drops, with SRL it's exactly the other way around.
Instinctively more natural, one ply deeper it is not, I understand
I think I will keep my elo formula.
BTW, the NICE tool is developed for engine tuning, that it also can produce a reasonable reliable rating list was an unexpected bonus.
I had no idea that you made the tools behind this public or that they were meant for tuning!
I’ve spent all week working on engine tuning using EPDs! I even made a thread asking for help getting quiescent positions. I wish I knew about your tool and datasets beforehand. NICE is a really cool project and really well executed. Thanks for making it public.