Arasan 10 - short time controls (John?)

[bob]

I have noticed that in the fastish 1+1 time controls I normally use, (one min on clock 1 sec added per move) that out of about 8000 games against each opponent, Arasan 10 will lose maybe 10 games on time. No other program loses even a single game, including Crafty which is playing almost 40,000 total per run.

I wanted to make a "quick check" test and thought I would try game/3sec time control, since I have played literally millions of test games at game/1sec with crafty vs crafty over the years. I use a time control of 0+0, but have my referee do just like xboard does and assume that the 0 base time really means 10 seconds. Except I am using 3. every program except arasan is doing OK here, but arasan is losing _many_ games on time control. In fact, it is only winning around 1% total, which is way less than at 1+1.

My referee sends the usual time/otim 300 (300 1/100ths of a second) for the initial time and everybody seems to play, but Arasan is stepping over that badly. Not on the first or second move, but the games rarely go any distance. Is this a known issue or is there a timing fix I should be using???

[Michael Sherwin]

Jon!

[swami]

I assume Jon only reads the General topic forum where he frequently takes part in topics where the test position is involved..

So moving this thread back to general topic may make sense? (Since this thread concerning the engine bug is more suitable in general topics sub forum anyway)

Can I have original poster's permission to do that?

[bob]

I assume Jon only reads the General topic forum where he frequently takes part in topics where the test position is involved..

So moving this thread back to general topic may make sense? (Since this thread concerning the engine bug is more suitable in general topics sub forum anyway)

Can I have original poster's permission to do that?

certainly...

[jdart]

Well, I do check this forum from time to time but not daily. I'm most always reachable via email at http://www.arasanchess.org/contact.shtml.

First, no it's not a known issue. At ICC blitz levels (it doesn't play bullet) it doesn't lose on time. But I don't routinely test Arasan at very fast time controls. It doesn't accept bullet matches on the server. So it is possible there is a problem.

I have also seen some latency issues with winboard/xboard - where there is a lag between the engine sending something and it getting processed - but it doesn't sound like that is the issue here.

I'll take a closer look at it. The timing logic is all in calctime.cpp if you want to see what it is doing.

--Jon

[bob]

Well, I do check this forum from time to time but not daily. I'm most always reachable via email at http://www.arasanchess.org/contact.shtml.

First, no it's not a known issue. At ICC blitz levels (it doesn't play bullet) it doesn't lose on time. But I don't routinely test Arasan at very fast time controls. It doesn't accept bullet matches on the server. So it is possible there is a problem.

I have also seen some latency issues with winboard/xboard - where there is a lag between the engine sending something and it getting processed - but it doesn't sound like that is the issue here.

I'll take a closer look at it. The timing logic is all in calctime.cpp if you want to see what it is doing.

--Jon

If you want to test a change, send it to me and I'll be happy to roll it in to my cluster test. I ran a 1+0 test last night to see how much quicker it runs, and in that time control, everybody lost a few games (few = 10 out of 8000 or so) except for arasan, where it was 857...

[opraus]

Hi,

I too was forced to remove Arasan as a sparring partner for same reason.

Could their be a problem here? [snipped from search.cpp]

Code: Select all

if (stats->elapsed_time > 100) {
         long denom = bkgrnd_search ? 200L : 500L;
         Time_Check_Interval = (int)(
         (denom*stats->num_nodes)/(36L*stats->elapsed_time));
}

My guess is that Jon is approximating a good poll rate. Is it possible it is too low?

-David

[jdart]

I think this is not the problem. There is some code in calctime.cpp that sets a minimum of 30 ms for a search. That is probably it but I am doing some testing to see.

[jdart]

I sent you a fix to try.

Frankly, though, I think game/3 seconds is pretty ridiculous. I have all kinds of depth-dependent code that may not get hit at this time period, plus it's not probably going to get the full benefit of filling up the hash tables and other caches, so the NPS will be really low.

Yes, it will play chess but it won't act like a normal search would, so comparing performance with other programs on this basis is questionable IMO.

[bob]

I sent you a fix to try.

Frankly, though, I think game/3 seconds is pretty ridiculous. I have all kinds of depth-dependent code that may not get hit at this time period, plus it's not probably going to get the full benefit of filling up the hash tables and other caches, so the NPS will be really low.

Yes, it will play chess but it won't act like a normal search would, so comparing performance with other programs on this basis is questionable IMO.

I am trying to find a quick way to measure improvements, not specifically measure the strength of each program. I agree it is very fast. But it is also a pretty nice facility to be able to make a change and play 2560 games per minute, or get a full 40K game result back in around 15 minutes or less. I always run longer tests to confirm results, but this is a pretty quick way to test several ideas without waiting days for all the results...

I am now trying 10s+.1s which is slower and the .1s increment should stop most flag problems since running out of time is harder. This takes about 1 hour to complete, which is still not bad, but not as good as the real quick 15 minute runs...