New Fischerle Version: 0.9.65

[RolandoFurioso]

Hi all,

I have just released a follow-up version of my chess engine: Fischerle 0.9.65. It is available for download at http://www.stuckardt.de/index.php/schac ... herle.html

The material evaluation has been enhanced and, in a certain regard, simplified. Moreover, some minor inaccuracies in the treatment of particular endgame types have been eliminated.

Fischerle 0.9.65 should play moderately stronger than its predecessor (Fischerle 0.9.60), the current score of which amounts to 2258 @ CCRL 40/40.

As its predecessor, Fischerle 0.9.65 features both a 64 bit and a 32 bit version, though it is highly recommended to run it in a 64 bit environment.

Have fun!
Roland

[Sylwy]

Thank you !

Already in test from two days .

SilvianR

[Graham Banks]

Excellent news!

Thanks Roland.

[Dariusz]

Good news!
Thx

[Sylwy]

http://talkchess.com/forum/viewtopic.ph ... 76&t=50060

SilvianR

[RolandoFurioso]

Hi Ruxy,

thanks for testing the new Fischerle version!

I think it is yet too early to judge upon its CCRL 40/4 strength, since it should be tested against much more opponents. There are always opponents against which an engine performs over or under average, so we will have to wait until a reasonable number of games have been played against a larger set of engines.

My (to a certain extent informed) guess is that the new version will score just a bit higher than the previous one - but let's see!

Roland

[RolandoFurioso]

http://talkchess.com/forum/viewtopic.ph ... 76&t=50060

SilvianR

Update - September 26, 2015:

Fischerle 0.9.65 64-bit @ CCRL 40/4: 2296, 543 games played - a plus of 69 compared to 2227 of its predecesssor Fischerle 0.9.60 64-bit.

Fischerle 0.9.65 64-bit @ CCRL 40/40: 2244, 104 games played - a minus of 14 compared to 2258 of its predecesssor Fischerle 0.9.60 64-bit.

Thus, there seems to be a considerable improvement at the faster time control, whereas there seems to be a slight deterioration at the slower time control.

Looking, however, at my own results, I lean to say that the current 40/4 score is still a bit too high (although - somewhat surprisingly - already 543 games have been played), whereas the current 40/40 score is yet too low. Since Fischerle employs a relatively knowledge-rich evaluation function, it typically scores higher under slower time controls; I consider it unlikely that this tendency has been reversed for the new version, given the relatively small changes.

So let's see how the rankings further develop for the next 200-300 games - in particular, regarding the CCRL-40/40 testings.

Roland

[RolandoFurioso]

Hi Roland,

Fischerle games on the list are from me. I have an i5 4690K CPU, my OS is 64-bit Windows 10 Home. Who knows, that may account for some of the differences if 40/40 games are played on a different kind of CPU.

Best regards,

Hi Gabor,

thanks for running Fischerle's 40/4 testings - I already had been pretty sure that you were running these particular tests

I definitely share your thoughts that the system environment might be another factor here. Specifically, I think that the particular combination of operating system and Java Runtime Environment (JRE) version could account for non-negligible performance differences.

Put differently: in case of Java engines, the code is not directly executed by the system, but indirectly by the Java Virtual Machine, which might perform differently across varying environments. That is, if one wants to proceed very exactly, one could employ a different calibration. In effect, this would imply that different time controls might have to be assigned for non-Java vs. Java engines - which, of course, seems to be hardly feasible.

Regarding the currently running 40/40 gauntlet, there are some indicators that the JVM running Fischerle is, to a certain extent, undersupplied with memory - that is, every time Fischerle starts to think about a new position (in the first couple of iterative deepening iterations), it takes some time until it reaches its usual processing speed (as if there were some sort of paging (access of slow virtual memory) occurring, which sould not happen). This might well account for a rating difference of 30 to 50 Elo or so. The results that you obtained on your system at the 40/4 time control are more in line with my own results.

Anyway - just let's see how Fischere will perform in the upcoming regular CCRL-40/40 tournaments.

Best,
Roland




a specific calibration -

[RolandoFurioso]

Hi Roland,

Some more games have been played under 40/4 and results are now closer to your expectation.

I have seen - thanks! Yet a pretty nice gain in CCRL-40/4 Elo anyway!

At CCRL-40/40 as well, Fischerle 0.9.65 has come nearer to the score that I expected.