Difference between Computer and humans chess players

[RubiChess]

Hi Andreas,

self-explanatory.
(... Frank describing how the statistics are created for the x.th time...)

I understand very well how the tables are created.
What I don't understand is: What do you believe to learn from the tables?
Again you have nothing said about my arguments regarding table 4 and Fire.

In all of the three ways Dragon by Komodo on on rank 1 for part 3 (passage into endgame).

I come with the same argument that I used for phase 4 and Fire:
Stockfish stronger than Komodo winning more winnable games in phase 2 while Komodo needs phase 3 for some more of them.
So "K is better than SF in passage into endgame" is not what you can learn from this statistics.

Give the engines a book with 100% positions all in "passage into endgame" and play a tourney. If this tourney shows the same ranking, I may believe in your conclusions.

Regards, Andreas

[Frank Quisinsky]

Hi Andreas,

learn from the table you wrote ...

I have interest to know ...
Which engines have which weaknesses and strong-points!
I can't see that if I looking on boring rating-lists.

So I am thinking:
Frank, do it in self work if I can't find it.

To table 4 and Fire I wrote a bit in my posting before.

Again to the examples to Stockfish / Komodo:
If Stockfish have a clear advantage in the first gaming phase (not only with the double quantity of fast won games in FCP Tourney-2022) and the different is around 20 Elo only, Komodo must have a strong point in comparing with Stockfish. That is logicial! If both engines do the same, the statistics are the same. And the strong point is gaming phase 3. The strong point is a long time again and again gaming phase 3 if I compare Stockfish with Komodo ... now Dragon by Komodo. That is not new for me.

Topic book:
Not a topic for myself.

FEOBOS opening is checked more as 10 times with more as 10 engines with long-time analyzes for each position.
All the engines have to play with the same opening book.

A great deal for the programmers!!
Because most of short-lost games must be blunders and not a problem from the opening book.
So the topic book isn't a topic for me.
But I will make FEOBOS better after the FCP Tourney-KI with 1 move more ...

Three moves after ECO codes building at the moment
Four moves after ECO code building with the next project

For this new project I need a lot of games and a good mix between Computerchess, correspondence chess and human chess on high level. One of the reasons for my FCP Tourney-Ki I started.

Best
Frank

[Frank Quisinsky]

Andreas,

I believe that for you is all a bit "magic".
Should be clear ... RubiChess produced very equal stats to all the gaming phases.
Move-average is great, king safty with many pieces is great ...

Prime example for ... a program I can see nothing on weaknesses with statistics for the strength your program produced.
All is very balanced!

So, you wonder as programmer of RubiChess.
Think so ...



Best
Frank

[RubiChess]

Hi Andreas,

learn from the table you wrote ...
I have interest to know ...

Pardon? I wrote no table. Just telling you that in my oppinion you take the wrong conclusions from your tables.

Which engines have which weaknesses and strong-points!
I can't see that if I looking on boring rating-lists.

I agree.

To table 4 and Fire I wrote a bit in my posting before.

But nothing about my argument "Fire just not strong enough to finish the winnable games before". The table says that it needed to play 303 "endgames" while Stockfish only played 249 which is a difference of 54 games ~ 21% while the score difference (which makes the ranking) is only 3%.
Lets assume some Stockfish developer is interested to improve what you call endgame strength and he writes a patch that delays games that will be won by some shuffling (to reach more moves for your move-based splitting) or by exchanging useless material (for your piece-based splitting).
Lets assume that after this patch Stockfish also plays 303 games in the endgame stage. It will reach additional score of 54 in your endgame table now with 212,5 points total = 70% score. 1st place now for Stockfish with a big gap to Fire now on 2nd. But has the patch improved endgame play? Of course not!
Traue keiner Statistik, die du nicht selbst gefälscht hast!

Topic book:
Not a topic for myself.
FEOBOS opening is checked more as 10 times with more as 10 engines with long-time analyzes for each position.

No doubt that your FEOBOS book is a good one. But it is not useful for testing endgame facilities.
Because (too) many games are finished long before reaching the endgame phase.

All the engines have to play with the same opening book.

Of course. The book only gives the starting positions of the matches that are played twice with reversed colors.
But if you really want to test endgame facilities you should use a book with positions starting at least near the endgame.

Enough from my side now.

Regards, Andreas

[Frank Quisinsky]

Hi Andreas,

table-4 (endgame) and Fire:
Indeed good arguments!

Fire won 98 / 303 games = 32,34%
Stockfish won 61 / 249 games = 24,50%
Koivisto won 153 / 474 games = 32,28% ... I wrote it before ... a long time on 1 during FCP Tourney-2022 is still running.

For Ordo the situation is quite clearly.

But I can follow what you wrote here.
It's one statistic only Andreas!

You wrote:
Fire just not strong enough to finish the winnable games before!

Yes, that is the reason that in the gaming phase before Fire isn't one place 1.

Now, if Stockfish produced more games for phase 4 absolutely possible that more win games will be the final result. But in this case Stockfish is not on place 1 in phase 1 or 2. With other words ... the view must be in combination of all the gaming phases.

The higher the gaming phase that more complicated are statistics.
Very easy to do all that with the first gaming phase, not easy with the other gaming phases.

An other statistics for endgames Fire is on place 2.
I am looking to eval on move 80 for games go to move 100-300.
Search the games the eval is well adjusted.
And now I am looking how many win games Fire produced and how many win games Stockfish produced.

All the result is a new database and run Ordo again!

Here Stockfish have a very, very small advantage to Fire and Komodo.
Again, Elo in endgames ... many programs have around the same level.
Not really important Fire on 1, Stockfish on 1 or others on rank 1.
Differences are to small.

Much more interesting is to look in gaming phase 3.

You wrote in German language:
Traue keiner Statistik, die du nicht selbst gefälscht hast!

Da ist auch etwas dran!
Man kann vieles positiv darstellen was negativ ist.
Nur das macht jetzt für mich und meinem Zweck keinen Sinn.

Again to the book:
Lines to around 74 of 500 ECO codes goes very fast in endgames.
Such a statistic I made during FEOBOS is still running.

I am not a fan of most of the lines and working with a small trick.
All in (all ECO codes should be play) ... but

I gave C00-C99 a higher priority in the FEOBOS version I am using.
With the result that more E00-E99 games can be produced also.
With lesser D00-D99 lines I produced not so many of games, goes to fast in the endgame.

2-3% not more but a small trick I used!



Best
Frank

[RubiChess]

Just one last remark that I can't resist...

Now, if Stockfish produced more games for phase 4 absolutely possible that more win games will be the final result. But in this case Stockfish is not on place 1 in phase 1 or 2. With other words ... the view must be in combination of all the gaming phases.

Completely true. And the best combination of all gaming phases is... in the boring ranking lists.

[Frank Quisinsky]

Andreas,

Let us make an experiment with RubiChess if you like?!

Give Ruby again more pruning.
Made the pawns more aggressively.
Give the program more aggressiveness with pawn moves again opponents castling.
Give open lines more attention.
Hold the bishops and knights in the game.
Hold the queen in the game.
And so one ...

All what you have in your brain to create a very aggressive version of RubiChess.
Give me the "Experimental" version for a run in my KI-Tourney.
And let us look on statistics after ...

That is my main point.
Engines produced aggressivess on a very experimental and high level are for humans very interesting and often unlogical things produced great results.

Can be very interesting in case of RubiChess because at the moment all seems to be fine adjusted.
But what will be the result with longer time-controls if you do that?

Loud thinking ...
Have a nice evening!

Best
Frank

[Uri Blass]

I suggest the following to test best engine in every type of endgame by the following way.

1)Take a lot of engine-engine games
2)Find the first position when there is some type of endgame
3)Use all these positions as the starting positions for both sides

I guess usually the top engines are going to get 1-1 in simple pawns endgames and only in rare cases we are going to see a pawn endgame when engine A wins against engine B with white but draw with black.
Maybe we are going to get less draws in more complex endgames and it is interesting to see what are the endgames when we get relatively more decisive results.

Type of endgames ordered by material to test when I do not include pawns with the following material values
queen=10,rook=5,bishop=3.3,knight=3.2
1)pawn endgames(total non pawns 0)
2)Pawns against a knight and pawns(total non pawns 3.2)
3)Pawns against a bishop and pawns(total non pawns 3.3)
4)Pawns against a rook and pawns(total non pawns 5)
5)A knight and pawns against a knight and pawns(total non pawns 6.4)
6)2 knights and pawns against pawns(total non pawns 6.4)
7)A knight and pawns against a bishop and pawns(total non pawns 6.5)
8)A bishop and a knight and pawns against pawns(total non pawns 6.5)
9)A bishop and pawns against a bishop and pawns(total non pawns 6.6)
10)2 bishops and pawns against pawns(total non pawns 6.6)
11)A rook and pawns against a knight and pawns(total non pawns 8.2)
12)A rook and knight and pawns against pawns(total non pawns 8.2)
13)A rook and pawns against a bishop and pawns(total non pawns 8.3)
14)A rook and a bishop and pawns against pawns(total non pawns 8.3)
15)2 knights and pawns against one knight and pawns(total non pawns 9.6)
16)3 knights and pawns against pawns(total non pawns 9.6 and I guess practically never happen)
17)2 knights and pawns against one bishop and pawns(total non pawns 9.7)
18)A knight and a bishop and pawns against a knight with pawns(total non pawns 9.7)
19)2 knights and a bishop and pawns against pawns(total non pawns 9.7)
20)2 bishops and pawns against a knight with pawns(total non pawns 9.8)
21)A Bishop and a knight and pawns against a bishop with pawns(total non pawns 9.8)
22)2 bishops and a knight with pawns against pawns(total non pawns 9.8)
23)2 bishops and pawns against a bishop and pawns(total non pawns 9.9)
24)3 bishops and pawns against pawns(total non pawns 9.9 and I guess practically never happens)
25)queen against pawns(total non pawns 10)
26)A rook and pawns against A rook and pawns(total non pawns 10)

[Frank Quisinsky]

Completely true. And the best combination of all gaming phases is... in the boring ranking lists.

Maybe!



The reason I produced also a boring ranking list.

But I can better sleeping if I have more for looking.

The "Troja" interview with Stefan Meyer-Kahlen I made in the past with a picture from a full trash and the question ...
How many ideas goes after a short time in the trash?

A lot of ideas Frank ... his answer.

It is the same with statistics.
My trash is full with ideas I had in the past.
I lost many time ... but often I am thinking ... this new idea can be good.

Computerchess is for me more interesting if I try to do a bit with all the produced games I have.
Believe me I make much more stats as I public on my site. But with the most interesting stats (pawn structures) I am not sure. I try to find helps in chess books about the topics I like.

Best
Frank

[Frank Quisinsky]

Hi Uri,

I read your name here so long I am reading TalkChess.
You have a lot of knowledge!

Your opinion to the question: Endgame
Will be very interesting.

I am thinking that in pure endgames the different between strongest engines is very small. Pawn structures is the most important topic here. Furthermore, in the past we are looking not enough on situations with different constellation of figures on board.

Example: 2 bishops with a rook and two pawns vs. knight and two rooks and two pawns and so one.
Often I am thinking that more strength comes with neural network in for humans unlogical situations.

The question for you:
Do you think that in pure endgames (around 10-12 pieces on board) TOP Engines produced big differents in playing strength?

I agree with your posting!

Best
Frank