Release : Chess System Tal EXTREME

[Rebel]

From : https://www.sp-cc.de/index.htm

Code: Select all

                         early           bad  avg.win 
Rank  EAS-Score  sacs    sacs   shorts  draws  moves  Engine/player 
---------------------------------------------------------------------------
   1    818536  55.32%  43.85%  53.48%  06.26%   63   CS Tal Extreme 
   2    750655  57.26%  45.49%  44.36%  05.00%   66   Patricia 5 xtremeEAS  
   3    748559  52.40%  42.58%  49.75%  08.28%   65   Rebel Extreme 1.1  
   4    712625  51.61%  43.75%  41.43%  05.49%   66   Patricia 3.1 avx2  
   5    653831  50.65%  39.79%  41.13%  07.20%   68   CSTal 2.1 EAS  
   6    643722  50.79%  41.53%  42.61%  05.85%   67   Patricia 5.0 avx512  
   7    638648  52.20%  47.83%  39.07%  05.85%   67   Cerberus 2722145081r9b  
   8    555597  46.96%  40.00%  34.95%  03.46%   70   Patricia 4.0 avx2  
   9    526843  43.56%  41.14%  33.89%  04.21%   68   Cerberus 21124081r81  
  10    428179  37.88%  38.53%  28.30%  08.79%   75   Rebel Extreme avx2 

[Peter Berger]

Not only Rebel EAS 3.0 scored 150.000 style points more than 2.0, surprisingly it also is considerable stronger than the Patricia's. And thus I changed my mind.

What does "And thus I changed my mind." mean here?

Personally I would be very much interested in engines that find a good compromise between EAS score and playing strength.

My personal very first impression with Chess System Tal extreme ( still way too little games so far here) is a bit similar to that of Patricia 5.0 extreme. It feels a little bit overtuned when looking at the games. At times they both seem to make pretty random sacs that are just good for scoring some additional EAS points but don't feel too natural. ELO might be useful for balance here.

If you pit them against older engines, their main strength is that they can outsearch them, even if they took the saccing too far.

I would be interested to see some EAS values for the later Junior engines. If style programming becomes a thing, they should be remembered as being at the top strength-wise while regularly doing real positional sacs at the same time.

[peter]

Not only Rebel EAS 3.0 scored 150.000 style points more than 2.0, surprisingly it also is considerable stronger than the Patricia's. And thus I changed my mind.

What does "And thus I changed my mind." mean here?

Personally I would be very much interested in engines that find a good compromise between EAS score and playing strength.

My personal very first impression with Chess System Tal extreme ( still way too little games so far here) is a bit similar to that of Patricia 5.0 extreme. It feels a little bit overtuned when looking at the games. At times they both seem to make pretty random sacs that are just good for scoring some additional EAS points but don't feel too natural. ELO might be useful for balance here.

If you pit them against older engines, their main strength is that they can outsearch them, even if they took the saccing too far.

I would be interested to see some EAS values for the later Junior engines. If style programming becomes a thing, they should be remembered as being at the top strength-wise while regularly doing real positional sacs at the same time.

+1

[Rebel]

Not only Rebel EAS 3.0 scored 150.000 style points more than 2.0, surprisingly it also is considerable stronger than the Patricia's. And thus I changed my mind.

What does "And thus I changed my mind." mean here?

That it would make sense to release it.

Personally I would be very much interested in engines that find a good compromise between EAS score and playing strength.

Cool.

My personal very first impression with Chess System Tal extreme ( still way too little games so far here) is a bit similar to that of Patricia 5.0 extreme. It feels a little bit overtuned when looking at the games. At times they both seem to make pretty random sacs that are just good for scoring some additional EAS points but don't feel too natural. ELO might be useful for balance here.

It's definitely over-tuned else no cigar, but it need to win its games else no style points.

If you pit them against older engines, their main strength is that they can outsearch them, even if they took the saccing too far.

I would be interested to see some EAS values for the later Junior engines. If style programming becomes a thing, they should be remembered as being at the top strength-wise while regularly doing real positional sacs at the same time.

It can be done for every engine, playing self games, making a neural net and put it under the CS-Tal / Rebel. In the past I have done it for Rodent.

[peter]

Not only Rebel EAS 3.0 scored 150.000 style points more than 2.0, surprisingly it also is considerable stronger than the Patricia's. And thus I changed my mind.

What does "And thus I changed my mind." mean here?

That it would make sense to release it.

+1

[Peter Berger]

I would be interested to see some EAS values for the later Junior engines. If style programming becomes a thing, they should be remembered as being at the top strength-wise while regularly doing real positional sacs at the same time.

It can be done for every engine, playing self games, making a neural net and put it under the CS-Tal / Rebel. In the past I have done it for Rodent.

Why is this necessary? Is Junior too weak to win enough games to score it?

[GenoM]

Not only Rebel EAS 3.0 scored 150.000 style points more than 2.0, surprisingly it also is considerable stronger than the Patricia's. And thus I changed my mind.

[...]
My personal very first impression with Chess System Tal extreme ( still way too little games so far here) is a bit similar to that of Patricia 5.0 extreme. It feels a little bit overtuned when looking at the games. At times they both seem to make pretty random sacs that are just good for scoring some additional EAS points but don't feel too natural. ELO might be useful for balance here.

[...]

+1

[Rebel]

I would be interested to see some EAS values for the later Junior engines. If style programming becomes a thing, they should be remembered as being at the top strength-wise while regularly doing real positional sacs at the same time.

It can be done for every engine, playing self games, making a neural net and put it under the CS-Tal / Rebel. In the past I have done it for Rodent.

Why is this necessary? Is Junior too weak to win enough games to score it?

Junior itself does not have NN support, to create a NN for Junior (or any other engine) you at minimum play 20-30 million super fast self play games and create about 2.5-3 billion positions for a neural net. It creates a sort of fingerprint of the playing style of Junior, or any other engine. And the bonus (provided you have done things right) is an elo gain of hundreds of points. And then of course the created neural net needs a host, can be any NN engine.

[Peter Berger]

Junior itself does not have NN support, to create a NN for Junior (or any other engine) you at minimum play 20-30 million super fast self play games and create about 2.5-3 billion positions for a neural net. It creates a sort of fingerprint of the playing style of Junior, or any other engine. And the bonus (provided you have done things right) is an elo gain of hundreds of points. And then of course the created neural net needs a host, can be any NN engine.

I understand that, and it would be +very+ cool to see such an engine. But it is unlikely, as it would demand the collaboration of Amir Ban, who has retired.

I asked for less. Couldn't we get an EAS score for one of the latest Juniors by someone who has a setup to do a few thousand games easily? They shouldn't be more than 200 points weaker than CS Tal extreme if my calculation ability is not way off, should they?

Peter

[Rebel]

Junior itself does not have NN support, to create a NN for Junior (or any other engine) you at minimum play 20-30 million super fast self play games and create about 2.5-3 billion positions for a neural net. It creates a sort of fingerprint of the playing style of Junior, or any other engine. And the bonus (provided you have done things right) is an elo gain of hundreds of points. And then of course the created neural net needs a host, can be any NN engine.

I understand that, and it would be +very+ cool to see such an engine. But it is unlikely, as it would demand the collaboration of Amir Ban, who has retired.

I asked for less. Couldn't we get an EAS score for one of the latest Juniors by someone who has a setup to do a few thousand games easily? They shouldn't be more than 200 points weaker than CS Tal extreme if my calculation ability is not way off, should they?

Peter

I don't have a Junior engine, but if you have I can give you a setup to play a few thousand Junior games yourself.