Winter is Coming (Amoeba, Beef, Combusk, Demoli, Fab, Halogen, Koivi, Lc0, Marvin, Minic, Nemori, Orion, Pedone, Winter)

[Frank Quisinsky]

Hi Martin,

I am thinking that you build such an opinion.
And all is fine for me.

Most of people are searching again and again the newest developments.
Yes, persons like to read TalkChess!

How many persons from the chess computer community read TalkChess?

Example:
From 2018 - start of 2020 I purchase 70 chess computers.
Middle of 2020 I am selling 70 chess computers.

I like to play a tournament with all the older computers and have here a lot of fun.
Not a topic I will do up-to end of life.

Now I am looking in new engines and organize again tournaments.
After all this I have interest to install all the engines from Winboard times, or Arena times.
ZChess vs. Phalanx will be a big topic for me in the future.
Or Zarkov vs. Crafty!

I hope John have all the older Zarkov sources.
No, no ...



The older engines / chess computers are never forgotten for people like computer chess.

The older chess computers have a very big community!!
Really nice people with great ideas in tuning chess computers.

My chess friend installed for two weaks 16 older Winboard engines from the beginning of the aera.
The lucky guy is now looking in the matches and after all I know with a lot of fun.
Never the guy read TalkChess ...

To your list of engines:
I like all of the engines, not Rybka

Isa?
An order for today, completly unknown for me.
I will find out that.

I am sure you will have fun with chess programming and each own way is something special!!

Cheng4 will be start in 7 days for FCP Qualify Tourney-2020 with 43 other engines.
You can be sure that I observe the games from your new Cheng version and you can be sure I am looking not on the results ... more on the style.



Best for you and thanks again for your engine!
Frank

With or without a new version.
Cheng 0.39 was at first on the list of participate engines of this qualify tourney!
OK, now with Cheng 0.40

[Angle]

I believe that style comes from search and eval inaccuracies,

I suppose you are somewhat ironic, but if take it seriously, I absolutely disagree! I am strongly convinced that even among perfect chess games (that don't contain inaccuracies at all) there exist representatives of all possible play styles: tactical, positional, attacking, defensive, maneuvering etc. Despite the fact that the results of all perfect games are the same, the ways to achieve this result may be fundamentally different which determines different play styles and approaches to chess. Moreover, I strongly believe that only within the perfect play the style can be expressed in its pure and complete form. If someone's play is full of inaccuracies and blunders then ir can't represent any style (unless you interpret a chaotic generation of random moves as a ''style''). Thus, all ''search and eval inaccuracies'' should be treated not as a source of style, but as a source of ''white noise'' within the style. Obviously, any white noise can't create a style as well as radio interference cannot create a style of music on the radio, but can only spoil it. However, some sources of noise can generate it with certain patterns and individual traits, but this forms not the play style. but what can be called the ''error style'' which is by no means the same as play style.

after all, eval features are tuned automatically.

Tuning is just weights balancing, i,e. the last and technical stage of development. What really matters is features themselves, the components of eval and the heuristics of search which are individual for each engine and form its play style.

[mar]

Actually I was serious
I believe that most of the "style" (90+%) is determined by static evaluation.
I wonder if there's really that much diversity among recent NNUE-based engines, assuming they use the same weights. Then the only difference would be the search.
But it clearly shows us how much potential actually lies in static evaluation.
Also Orion NNUE did incredibly well in your tournament and even though handicapped by utilizing a single core, it finished in the 3rd place.

It would be interesting to try and tune eval weights based on human games, let's say a Capablanca personality. I doubt 600 games is enough but it might be a fun experiment.

[Angle]

Actually I was serious
I believe that most of the "style" (90+%) is determined by static evaluation.
I wonder if there's really that much diversity among recent NNUE-based engines, assuming they use the same weights. Then the only difference would be the search.
But it clearly shows us how much potential actually lies in static evaluation.
Also Orion NNUE did incredibly well in your tournament and even though handicapped by utilizing a single core, it finished in the 3rd place.

It would be interesting to try and tune eval weights based on human games, let's say a Capablanca personality. I doubt 600 games is enough but it might be a fun experiment.

This is very close to my own position. I don't have much interest in engines using SF NNUE as their *official eval* (such as BBC or Mayhem). I think that having a bunch of impaired Stockfish eval-clones at the top of rating lists should be a real deadlock. In my *official divisions* I am going to use only NNUE engines that are trained on original data and/or using original framework. I will include engines using SF NNUE only to my *fun tournaments* in order to see how much elo points SF NNUE can gain to a certain engine or to compare them with the versions of the same engine using the original/novel neural nets. Including Orion NNUE to this tourney is one of the examples. However, I don't include Orion NNUE to any of my official divisions where only Orion 0.7 (without NNUE) will play.

I believe that the problem of diversification NN's (evals) which you also mentioned should become the key task in the engine's development for the next few months (years?). Not everyone understand but this is not only the problem of engine's originality and style but also the problem of gaining additional strength. If all strong engines would use (almost) the same eval then they would become blind in the same way and they would have no chance to reval this blindness and panish it by playing games to each other. In order to expose this blindness and to heal it one need to use NN's with completely different eval properties and priorities. Thus, we need to create a great variety of foundamentally different NN''s (evals) in order to get stronger engines. Ideally, we would come to algorithms/software for auto-diversifying (or human-managed-diversifying) of the nets training process and for auto-selection the best features of individual networks within a large collection of competing NN's. Here I anticipate a peculiar application of genetic algorithms of some kind which would generate/evolve/interbreed/select/inprove large populations of neural networks having essentially different ptoprties.