Back to ACT4 ... Some analysis

I ran the following engines:

Crystal-x64-mingw
Fat-fritz2-x86-64-avx2
Fat-fritz2-x86-64-avx2-NH
Komodo-13.2.5-64bit
Lc0-68002
Lc0-512
Lc0-alt
Stockfish-x64-avx2
SugaR AI ICCF 1.90 avx2

against the ACT4.EPD test file. I then analyzed the results with this program:
https://drive.google.com/file/d/1gytB_d7IppVhN8eFvF15rk3DYr9b1gCD/view?usp=sharing

which consumes the Arena log file and the EPD test file and creates a report.

The spreadsheet in this archive:
https://drive.google.com/file/d/1E-ULffqyE1SAHb3TrGYp50q2Fkq7N3qq/view?usp=sharing

contains a summary of the analysis, ranking the problems in difficulty. These are the hardest problems:

Code:

PROBLEM   SCORE   SOLVERS
35   549   0
39   549   0
66   549   0
49   508   1
1   498   1

Problems 35, 39, and 66 had no solvers with 62 threads, 32 GB hash, Full 6 man Syzygy and Nalimov EGTB files on ultra fast SSD at one minute analysis time.

These are the easiest problems:

Code:

PROBLEM   SCORE   SOLVERS
77   141   9
44   92   9
85   87   9
9   62   9
42   43   9
46   43   9
65   19   9
8   2   9

Problems 8,9,42,44,46,65,77,85 were solved by all 9 engines given one minutes time. Problem 77 consumed a total of 141 seconds to solve, by all the engines combined and problem 8 consumed a total of 2 seconds to solve, by all the engines combined.

So this one is the easiest problem {by far} in the set:
8/b4kp1/2p1p3/5pB1/p1Pp4/1pK1P3/5P2/3N4 w - - bm exd4; id "ACT4.008";

And these are the hardest:
2k5/Q1np1N2/2p2Bpp/1p1P4/pPP1p2P/P5q1/8/1K6 w - - bm Qxc7+; id "ACT4.035";
8/8/2pppN2/p1q2kpP/p4p2/P2P1PP1/P1P1NP2/1K6 w - - bm Nd7; id "ACT4.039";
2q1k1b1/7p/4p1r1/Pp1pP2Q/1P1P2PB/2P5/1KP5/8 w - - bm Qxg6+; id "ACT4.066";

All of these problems have been painstakingly analyzed as correct with the last one consuming at least a hundred hours of computer analysis all by itself. There was an internet video dedicated to this problem, but I am no longer able to find it. You can look it up on YACPDB:
https://www.yacpdb.org/#static/home

I should mention that there is a problem with the LC0 analysis:
I ran the analysis with too many threads (62). That is not good for LC0.
So I am going to re-run the three LC0 trials with three threads instead of (gulp) 62.
I have two 2080 Super GPUs in my machine.

Here is something I did not expect. The impact of way too many threads is negative but very small

62 CPU threads
Analyzing engine: Lc0-68002
48 of 100 matching moves

Analyzing engine: Lc0-512
29 of 100 matching moves

Analyzing engine: Lc0-alt
53 of 100 matching moves

Three CPU threads
Analyzing engine: Lc0-68002
49 of 100 matching moves

Analyzing engine: Lc0-512
32 of 100 matching moves

Analyzing engine: Lc0-alt
54 of 100 matching moves