|
|
(1) Posted by Jakob Leck [Sunday, Jun 2, 2024 22:48] |
Testing long helpmates with lots of material When computer-testing long helpmates, there are different categories: Problems with not too much material, very mobile pieces, but limited possibilities for mating positions, e.g. P0510416. These are quickly solved by the intelligent mode in popeye. Then there are problems with a lot of material, but where the somehow restricted position allows for quick solving in intelligent mode, e.g. P0500863. Also there are positions with such little mobility that, although very long, they can easily be solved brute force, e.g. P0502947.
And then there are problems that do not exactly fit those categories, like P1376204, where computer testing in intelligent mode in popeye takes quite long (but, of course, is still preferable to brute force).
Can any of you recommend faster methods of testing such problems? (Faster by how much?) |
|
(2) Posted by Olaf Jenkner [Monday, Jun 3, 2024 22:05] |
I don't know what "quite long" means. The full test took Gustav 75 minutes (3333 MB hash). Gustav doesn't have an intelligent mode for helpmates.
It is a "half intelligent mode". If white has only one piece as in P1376204, the last three moves work with intelligent mode,
otherwise the last two moves. Gustav works without any optimization if a white pawn is on the 7th rank or of a white queen is on board. |
|
(3) Posted by Jakob Leck [Tuesday, Jun 4, 2024 22:13] |
What I mean by "quite long" is: On my laptop P0510416 is solved in seconds, the other two mentioned problems in a few minutes, but P1376204 takes well over an hour (I aborted it). Which is an inconvenience, but I'm trying to compose a h#7 with seemingly similar properties as far as computer-testing is concerned, where even the h#6 after the second two half-moves takes over an hour to test, the h#6,5 after the first half-move was still going when I aborted after half a day and I haven't tried the full 7-mover yet, but I'm looking for a significantly quicker method of checking it, if there is one. |
|
(4) Posted by Marko Ylijoki [Wednesday, Jun 5, 2024 04:27] |
Analyzer is significantly faster than Popeye.
https://helpman.komtera.lt |
|
(5) Posted by Jakob Leck [Wednesday, Jun 5, 2024 05:44] |
I guess it draws the solution from a database, right? Slightly alter the position (e.g. wB from d5 to c6 in my double Schnoebelen) and it actually starts to try solving, but says "Too long solving" after about 5 minutes. |
|
(6) Posted by Marko Ylijoki [Wednesday, Jun 5, 2024 06:39] |
Yes, sometimes such a small change seems to increase the evaluation time significantly. If wB was on e4, it would take less than 1.5 minutes. Maybe Viktoras can explain why. Anyway, I've been a very happy Analyzer user since my return. |
|
(7) Posted by Torsten Linß [Wednesday, Jun 5, 2024 11:03] |
If you cannot guess in advance which program (or which options) is the fastest, you should simply run them in parallel. Probably your notebook has 4 cores to facilitate this.
Torsten |
|
(8) Posted by Olaf Jenkner [Thursday, Jun 6, 2024 00:32] |
Alybadix might be the fastest program for this kind of helpmates. |
|
(9) Posted by Jakob Leck [Thursday, Jun 6, 2024 17:44] |
That's what I had in the back of my mind, having heard some years ago that Alybadix might be the fastest or among the fastest for such tasks. (Though I never tried to obtain or use it, thought I didn't need it and had no 32-bit Windows available anyway...) Are there any benchmarks available? (Like, comparing different programs for some problems on identical hardware?) Depending on the possible improvements it's likely better to invest in better hardware for running popeye than Alybadix.
In the meantime I've heeded Torsten's advice to try different methods in parallel. Thanks! |
|
(10) Posted by Viktoras Paliulionis [Monday, Jun 10, 2024 11:46] |
I have slightly improved the algorithm, and now Helpmate Analyzer solves P1376204 in less than 3.5 minutes. This is significantly faster than Popeye (1+ hour).
Previously, the Analyzer took longer to solve this problem but showed its solution immediately. This was because P1376204 was included as an example in the Glossary, and its solution had been cached to make it easier analysis for users. I have now removed it from the cache, so the Analyzer shows the actual solving time.
Typically, the Analyzer takes no more than 5 minutes to solve a helpmate of any length with one white bishop (even a few seconds are often enough), but there are exceptions. I would appreciate it if someone could provide examples with a single white bishop that the Analyzer cannot solve within 5 minutes. |
|
(11) Posted by Jakob Leck [Monday, Jun 10, 2024 17:11] |
This is absolutely amazing, Viktoras. Your analyzer took 1:35 m:s to solve a 2+16-piece h#6.5 that popeye worked on for more than half a day (a white bishop minimal).
If I come across positions that take longer than five minutes, I'll let you know. |
|
(12) Posted by Viktoras Paliulionis [Wednesday, Jun 19, 2024 21:31] |
There were a few bugs in the latest version of HA, but they have now been fixed. Additionally, the solver engine has been further optimized and now solves P1376204 in 1:11 min:s.
@Jakob, could you solve the same h#6.5 (2+16) and see how much the solving time changes? |
|
(13) Posted by Jakob Leck [Friday, Jun 21, 2024 17:07] |
For the h#6.5 the solving time was almost unchanged: 1:33.345 m:s
Also, indeed you seem to have fixed some bugs. I had an illegal 2+17 h#7 that had a single solution according to analyzer, but popeye showed cooks, and now analyzer finds cooks as well. :) |
|
No more posts |
MatPlus.Net Forum Internet and Computing Testing long helpmates with lots of material |