Glaurung 080420 evalaution

[ozziejoe]

I'll be testing glaurung08 against various opponents and comparing it to glaurung2-epsilon 5. Once i get a stable rating, I'll start tweaking parameters. any hint on what parameters are working so far?


Rybka test finished

conditions: 4 min 2 sec; core duo, 32 bit; engines use 2 threads if can; noomen 2006 test set


Rybka 2.3.2a mp versus glaurung 080420 ------- +37/-5/=18
rybka 2.3.2a mp versus glaurung2-epsilon/5----- +42/-3/=15

[Ryan Benitez]

Here is +5 and -5 to the king attack values to check on how well they are in tune with the rest of eval. I think the default value is very good. The next natural step is to try new values in the SafetyTable[] array. Based on experience I think it would be good to try with values climbing more quickly and peaking sooner. I could easily be wrong though.

Glaurung 080420-k105 : 2829 720 (+279,=234,-207), 55.0 %

Spike 1.2 Turin : 240 (+130,= 68,- 42), 68.3 %
Chess-64 07/11/11 : 240 (+ 97,= 80,- 63), 57.1 %
Fruit2.3.5p13 : 240 (+ 52,= 86,-102), 39.6 %

Glaurung 080420 : 2825 720 (+273,=238,-209), 54.4 %

Spike 1.2 Turin : 240 (+101,= 89,- 50), 60.6 %
Chess-64 07/11/11 : 240 (+104,= 76,- 60), 59.2 %
Fruit2.3.5p13 : 240 (+ 68,= 73,- 99), 43.5 %

Glaurung 080420-k95 : 2819 720 (+270,=231,-219), 53.5 %

Spike 1.2 Turin : 240 (+110,= 72,- 58), 60.8 %
Chess-64 07/11/11 : 240 (+113,= 77,- 50), 63.1 %
Fruit2.3.5p13 : 240 (+ 47,= 82,-111), 36.7 %

[Tord Romstad]

Hello,

Thanks to Ryan, Shaun and Tony for their results!

It looks like the new version might be strong enough to merit a public release, even in its current raw and untuned stage. I can always release a more polished version later, if I or somebody else can come up with significantly better settings.

At any rate, I'll probably spend at least another week or two running experiments and cleaning up the code before I release Glaurung 2.1.

Here is +5 and -5 to the king attack values to check on how well they are in tune with the rest of eval. I think the default value is very good. The next natural step is to try new values in the SafetyTable[] array. Based on experience I think it would be good to try with values climbing more quickly and peaking sooner. I could easily be wrong though.

Quite possible. The current development version actually contains several new UCI parameters for configuring how the king safety values climb. The SafetyTable[] array, from which the king safety scores are taken, is no longer a constant array, but is computed based on the values of UCI parameters.

Tord

[Ovyron]

At any rate, I'll probably spend at least another week or two running experiments and cleaning up the code before I release Glaurung 2.1.

Great news! I'm looking forward for next Glaurung release

[ozziejoe]

hi tord,

are their any changes in the newest glaurung that makes her (or him?) more aggressive.? My initial sense is that the new glaurung is a little less drawish than epsilon 5 (but I could be seeing things)

i think releasing it would be a good idea, especially with all the configurable parameters. Someone will discover a stronger personality...(someone with more computers than me)



best
J

[ozziejoe]

sorry tord...one more question....Do you have a description of what the parameters mean? I was wondering what aggressiveness (is this optimistic search?) versus cowardice (pessimistic search?) means, and of course I am interested in the other paramters.

By the way, i think the more parameters you can make configural, the more chance someone will have of finding a way to improve

best

J

[Dr.Wael Deeb]

hi tord,

are their any changes in the newest glaurung that makes her (or him?) more aggressive.? My initial sense is that the new glaurung is a little less drawish than epsilon 5 (but I could be seeing things)

i think releasing it would be a good idea, especially with all the configurable parameters. Someone will discover a stronger personality...(someone with more computers than me)



best
J

You need a lot of computers under your hands to find a strong personality,even though Glaurung has a hell lot of parameters....I have Glaurung 1.2.1 Witch MP which is 30-40 Elo stronger than the default and it's closing the gap between it and Glaurung 2.0.1 MP in my rating list

[Tord Romstad]

are their any changes in the newest glaurung that makes her (or him?) more aggressive.?

Yes. The main change compared to Glaurung 2.0.1 is the addition of king safety evaluation, which makes the engine noticably more aggressive.

sorry tord...one more question....Do you have a description of what the parameters mean? I was wondering what aggressiveness (is this optimistic search?) versus cowardice (pessimistic search?) means, and of course I am interested in the other paramters.

I don't remember exactly what parameters were present in the latest Windows binary I have (I've added a lot of new parameters since then), but I can answer about "aggressiveness" and "cowardice": These are not about the search, but about the king safety evaluation. "Aggressiveness" is the evaluation weight for attacks by the engine against the enemy king. "cowardice" is the evaluation weight for attacks by the opponent against the engine.

In other words, when "aggressiveness" is high and "cowardice" is low, Glaurung will like to play aggressive king side attacks, but won't care much about its own king. If "aggressiveness" is low and "cowardice" high, it will usually not attack, but it will be very careful to avoid king attacks by the opponent.

By the way, i think the more parameters you can make configural, the more chance someone will have of finding a way to improve

Perhaps, but too many parameters will be confusing to users. I now have close to 50 parameters, which is perhaps a bit too much. There are no less than 15 UCI parameters just for tuning the king safety evaluation. With careful tuning, it should be possible to find settings which are significantly better than the defaults, or which gives the wildest attacking engine since Chess System Tal, depending on what you prefer.

Tord

[ozziejoe]

it would be an interesting psychological experiment to see how the names of your parameters influence how people create personalities. e.g., now you have a parameter called "cowardice." Who amongst us really wants a more cowardly engine?

Maybe you should call aggressiveness something like "bullying" and you should call "cowardace" something like "disciplined restraint"

best
J

[Tord Romstad]

it would be an interesting psychological experiment to see how the names of your parameters influence how people create personalities. e.g., now you have a parameter called "cowardice." Who amongst us really wants a more cowardly engine?

Maybe you should call aggressiveness something like "bullying" and you should call "cowardace" something like "disciplined restraint"

Actually, the names "aggressiveness" and "cowardice" are taken from another well-known UCI chess engine. Can anyone guess which one?

If I had chosen names without external influence, I would probably have ended up with something dull like "King safety (engine)" and "King safety (opponent)". Perhaps this would have been less confusing to users.

Tord