Horvath Play Styles

spacious_mind · Post by **spacious_mind** » Sat Dec 27, 2014 10:04 pm

A little while ago I used my Test Game 1 (15th Century Master) to try out the different play styles that can be set in Gyula Horvath's chess computers.

Horvath's typically have 9 built-in play styles which range from ultra cautious to ultra aggressive. The standard out of the box play style is 5. You will always reset to play style 5 when you do the reset at the back of the computer or change batteries:

Here is what the manual says:

In addition to the 9 built-in play styles you can also create your own play styles from 7 configurable style settings ranging in value from 00 to 99 (100 settings per style setting). This configuration means that you have One Hundred Thousand Billion (100,000,000,000,000) possible configurations.

Here are the 7 configurable style settings:

If you read this carefully, then you should find this interesting as it explains a lot about Horvath's tendency to play a more passive game on your Horvath chess computer which the next chart will show in more details:

Now back to the 9 standard built-in style settings. If you change the style on a Horvath and then go and look inside the configurable style option, you will see the following parameter settings:

The standard out of the box parameter is style 5 which everyone plays with. If you notice the settings for style 5, every parameter is set to the value 10.

If you have read about FEATURE # 4 and FEATURE # 5 earlier, you would have seen it explained for both White and Black as "A big weighting will encourage a player's pieces to attack more than normal"!

Well since the standard out of the box setting has the values of 10 for every style with a low being 00 and a high being 99, it would stand to reason that the game style of Horvath computers would have a passive tendency!

Since I wanted to also rate the Horvath computers for a rating list, I also decided to compare the 9 different built-in styles to see how strong they are. I also used this opportunity to compare all the Horvath computers that I have, suspecting of course that each one of them would be identical.

The computers I tested were:

CXG SPHINX CONCERTO - 1992

CXG SPHINX LEGEND - 1992

EXCALIBUR LEGEND II - 1993

KRYPTON CHALLENGE - 1994

SYSTEMA CHALLENGE - 1994

Needless to say, Krypton Challenge and System Challenge played absolutely identical. So did CXG Sphinx Concerto and CXG Sphinx Legend. Therefore in the charts below they are lumped together.

The level I played was level 23 which is a 30 second per move fixed time level, which I think provides the best accuracy for this particular clone test. I played both white and black moves by every time taking back the computer move and reentering each move irrespective if the computer chose the same move as the test game. This kept everything constant and exactly the same for every move and every contestant.

HORVATH PLAY STYLES 9, 8 & 7

To my surprise after playing the first play style with these computers I saw that they are not all the same! In fact there are a few pattern and trends that can be seen very quickly.

White performs best at the highest aggressive setting. Black performs worst at the highest aggressive setting. Standard out of the Box style # 5 does seem to be the most balanced for both Black and White.

On the most passive settings White worsens whereas Black performs here better than at the most aggressive settings.

BEST SETTINGS TEST GAME 1

WHITE = STYLE 9
BLACK = STYLE 5
BEST OVERAL SETTING = STYLE 5

I think I am going to have to repeat this with Test Game 2 as well to get a better overall picture.

If you look at the charts above, you will also notice that there are 3 Horvath versions and they can be categorized by date:

Version 1 = 1992 = CXG Concerto & CXG Legend
Version 2 = 1993 = Excalibur Legend II
Version 3 = 1994 = Krypton/System Challenge

Legend 2 and Challenge most closely resemble each other, but there are very minor game differences which do repeat themselves.

I would have liked to include Krypton Regency to the above test but unfortunately my Regency is beginning to fail, not allowing me to take back moves or get to the feature screens. Other than that it plays fine.

Therefore missing to complete this Horvath Test are the following computers. If someone would volunteer that would be great!!!! as this would complete the Horvath comparison.

Krypton Regency
CXG Accolade
Excalibur Avenger.

If I were a betting man then I would bet that Accolade ends up being exactly the same as Concerto & Legend. Regency possibly exactly the same as Challenge. With Avenger I don't know but might be the same as Legend II or maybe Challenge. I doubt if it were the same as Concerto or Legend.

Best regards,

Nick

Reinfeld · Post by **Reinfeld** » Thu Jan 01, 2015 8:27 am

Wonderful stuff, Nick - again.

Based on this (admittedly limited) dataset, you've established that the ultra-aggressive style in this clone pod is stronger (as White) than the out-of-the-box setting of style 5 (which is better as Black). You are showing adjusted elo variations as wide as 444 points (1601 to 2075) in the same game.

Strikes me that there are really only 5 pre-programmed styles; 1, 2, 3 and 4 mirror 6, 7, 8 and 9 with opposite colors.

So now comes the real question: Is it possible to "tune" these Horvath models (using the programmable evaluation function) in a manner that yields a stronger player than any of the preprogrammed styles? That strikes me as the next level of your testing suite.

Just one for-instance outlier: what happens if you set all values to zero, or max them out? Will these be other versions of style 5?

You'll need more games for clarity, obviously. But the weighting of factors is pretty interesting from a chess standpoint. Picture a tuned style that values rooks on open files more than doubled pawns, for example. That would be my primitive instinct.

Looking forward to more results regards,

- R.

pr1uk · Post by **pr1uk** » Thu Jan 01, 2015 8:54 am

Fantastic work there Nick
Well Done

paulwise3 · Post by **paulwise3** » Fri Feb 13, 2015 11:15 pm

Nick,
Great test! I specially liked your diagram showing the parameter settings for the 9 playing styles. That is great material for analysis and searching for one's own personal playing style.
After some thinking I wonder if the 10-values for style 5 imply passive play (although it actually plays rather passive). It could well be that it is not the value, but the ratio between the parameter values that determines the priority. For instance, I suspect that setting all parameter values to 50 instead of 10, the playing style would be exactly the same, as long as they are all equal. But setting them all at 10 gives the user more easy flexiblity in assigning his own priorities. We should test that. Since a month I own a Concerto, so if I can find the time I will do some testing and let you know.

Anyway, this gives much food for thought!

Regards, Paul

spacious_mind · Post by **spacious_mind** » Mon Feb 16, 2015 4:34 pm

paulwise3 wrote:Nick,
Great test! I specially liked your diagram showing the parameter settings for the 9 playing styles. That is great material for analysis and searching for one's own personal playing style.
After some thinking I wonder if the 10-values for style 5 imply passive play (although it actually plays rather passive). It could well be that it is not the value, but the ratio between the parameter values that determines the priority. For instance, I suspect that setting all parameter values to 50 instead of 10, the playing style would be exactly the same, as long as they are all equal. But setting them all at 10 gives the user more easy flexiblity in assigning his own priorities. We should test that. Since a month I own a Concerto, so if I can find the time I will do some testing and let you know.

Anyway, this gives much food for thought!

Regards, Paul

Hi Paul,
Thanks, I am pleased you liked the test. Below is a chart where I had already tested your theory with everything set at 50 and 90. Also standard style 5 which has all values set at 10 is also included. The tests were done at level 23.

As you can see from the above chart, they differ slightly from each other, therefore your suspicion of setting all the styles equally ie 50 each as in style 5, has been busted

Best regards

paulwise3 · Post by **paulwise3** » Mon Feb 16, 2015 8:32 pm

Hi Nick,
Thanx for your reply. This result is very encouraging for me to do further testing. But the thing I am missing in this diagram is the resulting elo-score for the 50- and 90-parameter settings! Could you please add this? Are they better or worse than the standard 10?
(A while ago I downloaded your testsheet, but it needs a password to be really able to use it. I will post my moves in the appropriate thread shortly.)

Regards, Paul.

spacious_mind · Post by **spacious_mind** » Mon Feb 16, 2015 9:53 pm

paulwise3 wrote:Hi Nick,
Thanx for your reply. This result is very encouraging for me to do further testing. But the thing I am missing in this diagram is the resulting elo-score for the 50- and 90-parameter settings! Could you please add this? Are they better or worse than the standard 10?
(A while ago I downloaded your testsheet, but it needs a password to be really able to use it. I will post my moves in the appropriate thread shortly.)

Regards, Paul.

Hi Paul,

I have updated my previous Post to show also the scores. Standard Rating scored best. Settings all set to 90 however scored best as white.

Also please try and download the 5 tests again from the link below, I think all 5 should work now for you.

http://hiarcs.net/forums/viewtopic.php?t=6835&start=60

Please let me know if they work.

Thanks and regards

paulwise3 · Post by **paulwise3** » Fri Apr 03, 2015 10:40 pm

Hi Nick,
After an inspiring meeting with Hein Veldhuis and Luuk Hofman (they were very interested in this experiment, also they did an interesting discovery about another machine), I gave this style test a new try. I started with all features at weight 20, and that seems to be a good one. I have to mention though, that I tested at level 53 (30s average) as opposed to your tests with level 23 (30s fixed). Tonight I did testgames 1 to 3. In game 1 it scores better than all your tried settings (features at 10, 50 and 99).
In game 1 and three it also scores better then the Excalibur Grandmaster, game 2 it is weaker. So I am very curious how testgame 4 and 5 will come out!

Here the global results so far:

Code: Select all

Concerto_all_features_20
 test01 test02  test03  test04  test05	&#40;1..5&#41;/5
W-2332	2163	2323
B-2172	1755	2097
T-2252	1959	2210

Testing regards,
Paul

spacious_mind · Post by **spacious_mind** » Sat Apr 04, 2015 12:00 am

paulwise3 wrote:Hi Nick,
After an inspiring meeting with Hein Veldhuis and Luuk Hofman (they were very interested in this experiment, also they did an interesting discovery about another machine), I gave this style test a new try. I started with all features at weight 20, and that seems to be a good one. I have to mention though, that I tested at level 53 (30s average) as opposed to your tests with level 23 (30s fixed). Tonight I did testgames 1 to 3. In game 1 it scores better than all your tried settings (features at 10, 50 and 99).
In game 1 and three it also scores better then the Excalibur Grandmaster, game 2 it is weaker. So I am very curious how testgame 4 and 5 will come out!

Here the global results so far:
Code: Select all
Concerto_all_features_20
 test01 test02  test03  test04  test05	&#40;1..5&#41;/5
W-2332	2163	2323
B-2172	1755	2097
T-2252	1959	2210
Testing regards,
Paul

Hi Paul,

Nice to see you experimenting with the different Play Styles that you can create!

I am going to revisit the tests with Level 23 and add Krypton Regency. Based on the rating tests at level 53, Krypton Regency does not seem to be 100% the same as the others either. An observation with Regency is that it has a tendency to play moves too fast at level 23. It is rare that it goes the full 30 seconds when I am testing it.

ps. You should ask Luuk and Hein to come and post here from time to time. They have so much to offer us. I had bought quite a few computer from Luuk in the past (mainly from Tom's old collection), he is a gentleman.

Best regards

spacious_mind · Post by **spacious_mind** » Sat Apr 04, 2015 10:24 am

An observation with Regency is that it has a tendency to play moves too fast at level 23.

Sorry a typo, I meant to type too fast at level 53 (average time setting)

paulwise3 · Post by **paulwise3** » Sat Apr 04, 2015 11:33 am

Hi Nick,

Thanx for listing the score of the Carnelian II. Hope you have time to put the Grandmaster results that I sent you there too.
As for Hein and Luuk, they are working very hard on their own site documenting their machines and posting new items from Hein's overwhelming documentation. But I agree a post from them here would be nice.
The discovery I mentioned above was about the Chess King Counter Gambit. One thing was that it appears to play a lot stronger then is assumed here and there. The other thing was a difference with the pictures Mike Watters shows on his site: this one has an AC/DC connection, and Mike's machine has not. Interesting stuff for those looking for special ones.
To end with: the latest items of Hein's documentation are always shown on http://schaakcomputers.nl/ in the middle of the page under "Laatste items Heins database".

Best regards,
Paul.

paulwise3 · Post by **paulwise3** » Mon Apr 06, 2015 12:18 pm

spacious_mind wrote: Hi Paul,

Nice to see you experimenting with the different Play Styles that you can create!

Hi Nick,
Here are the completed tests for the Concerto with all feature weights at 20.
The average rating comes out about the same as with the normal style 5 settings, and also as the Grandmaster tests I did.
But it makes different moves, and I have the impression that it chooses more active moves. I will check this by playing some new games against the Sphinx Comet and the Europa.

Code: Select all

CXG Sphinx Concerto All feature weights 20.
 test01 test02  test03  test04  test05	&#40;1..5&#41;/5
W-2332	2163	2323	 1805	 2164 	 2157
B-2172	1755	2097	 2031	 1597 	 1930
T-2252	1959	2210	 1918	 1880 	 2044

Excalibur Grandmaster
 test01 test02  test03  test04  test05	&#40;1..5&#41;/5
W-2220	2341	2251	 2083	 1937 	 2166
B-2164	1881	1836	 2026	 1741 	 1930
T-2192	2111	2044	 2054	 1839 	 2048

If you like I can send you the details of each testgame.

Hoping for active play regards,
Paul.

paulwise3 · Post by **paulwise3** » Fri Apr 10, 2015 11:16 am

Reinfeld wrote: ...
Just one for-instance outlier: what happens if you set all values to zero, or max them out? Will these be other versions of style 5?
...

Last week I played a game with the Concerto with all values set to zero. Cannot say much about the playing style, but there was a funny thing: when I mated it, it kept thinking instead of signalling the mate! Tried it again by taking back and making the same move, it stayed thinking. Then again, take back and push the Move button, and see: it made the mate-ing move and signalled mate!?! Bad loser

.

@Nick: Looking at the feature settings for each style again, and also your testresults, one can conclude that level 1 is the most agressive for white. However, the manual states it is the most cautious level. That must be a misprint then I guess.

Questionmark regards,
Paul.

paulwise3 · Post by **paulwise3** » Mon Apr 13, 2015 8:44 pm

@Nick: Looking at the feature settings for each style again, and also your testresults, one can conclude that level 1 is the most agressive for white. However, the manual states it is the most cautious level. That must be a misprint then I guess.

Paul.

Correction, of course your tests show that level 9 is the most agressive. What misleads and confuses me, is that at style 1 the feature "AGRESSIVE PLAY BY WHITE'S PIECES" = 99, so one would think this one plays agressive, but apparently it is the opposite. And at style 9 the feature "AGRESSIVE PLAY BY BLACK'S PIECES" = 99, combined with "KING ATTACK BY PIECES"= 99.

Another observation: Yesterday I let the Concerto play against itself, once with style 1, and once with style 9. With style 1, black had a completely winning position, also materially speaking. But then it started to play strange moves, loosing all the material and the position. So it became a draw.
That's why in the second game, I decided to look at what evaluation values it gives a position while thinking. And there it was: a reasonably equal position only few moves after the opening was valued +4.49!? So I took back the move, set style to 5 (=normal) and pushed Move again. Now it showed a quite normal evaluation value of 0.04. Back to Style 9, and also a lot of the next moves still showed an advantage of +4.xx for white. Eventually White won this game.
This gives more food for study...

PM: I remember having read a review of some other machine with different playing styles. The advice for playing against other machines there was: Start playing Normal, switch immediate to Agressive when out of openingbook, and when you have clear advantage switch back to Normal. Having noticed the above, the same seems to go for the Legend/Concerto...

Headache regards,
Paul

spacious_mind · Post by **spacious_mind** » Mon Apr 13, 2015 10:42 pm

Hi Paul,

Well it sounds like you are having fun trying out the styles. It is not so easy understanding them. I was confused the other day when I played Regency at level 53 and it score was a lot less than what I showed for Legend/Concerto. Soon I will have to compare it closer to Challenge, because after maybe they are not exactly the same as previously suspected.

Best regards