Release: FEOBOS v20 Contempt 1-5 without move transp. avail.

[Frank Quisinsky]

Hi,

if you are looking in the database with move transposition you can see it ... (FEOBOS v20.1 download file).

Sorted with ECO codes, A00 at first ... each position to the same ECO code have the same length of moves.

Here you can see why different ECO codes have more as 12 moves (formed very late).

So I am working with 500 databases before 10 engines started the analysis ... 3 moves after ECO Code was formed.

---

Database with move transpositions: Sorted with ECO codes
Database without move transpositions: Sorted with ranking system

Forget to write that!

I read your message again!
I am very happy that you understand what we do with FEOBOS. My English is really not good.

Best
Frank

[Dann Corbit]

Hi,

10 Engines x 41.614 Positions x 60 Seconds = 299 Billions Nodes.
Intel 10 Threads, 6 Cores with 4.3 GHz.

More isn't possible for me! Double time = 2 years analyze time!!

Best
Frank

I have some big machines with lots of cores. I am going to analyze every book node, including the internal ones, to 36 plies or more.

[Frank Quisinsky]

Hi Dann,

that's better hardware I am using.

I made a test before I select the time for using!
For the first 1.000 of 26.146 positions (without upd.1).

30 seconds per move
60 seconds per move
180 seconds per move

Two questions:

1. How many other best / first moves engines found.
2. Different in evals (differents are very small, not interesting).

From 30 to 60 seconds I can see clearly better results, so I am using 60 seconds. At first, I am thinking 30 seconds are enough.

From 60 to 180 seconds lesser as 1% other moves engine Fire found. So, I am thinking time for analysis on my hardware with 60 seconds is good and I can start 10x 41.614 = 416.140 analysis.

Maybe in 3 years with 5x stronger Hardware and 100 Elo more for each of the engines we can produce a better result?!

Interesting is that 10 engine rejected in team work. With such a net you can be sure that the database is clear. Much more interesting as more time are many engines for Analysis. A Group of engines can found more as 2-3 engines with more time. That's the deal ... FEOBOS = teamwork by 10 engines!

Note:
6 Cores 10 Threads, 4.3Ghz with 60 seconds ...
No rating list will have such conditions in the next 3 years. Analysis will have the Quality important for rating list producer.

For us is the current result not more as a current rating list. Much more important is the Excel development because this work can be used again and again.

The Excel file we produced, with more as 11 million of Formulas, have very new and innovative idea. That was the most and a very hard work.

Between Klaus Wlotzka and myself 1.700 emails switched. Often, we solved problems 03:00 in the night. Most work we had with the ranking system.

A very hard work we do here in the last year. Much more work as in times I switched such development Information with Martin Blume for Arena.

Best
Frank

But your Hardware is great ...
You can see in KECR Excel that Houdini and Andscacs are most interesting for opening analysis. We can learn a lot to engine if we put over 41.146 positions a template. So many unknown details to engines, ECO codes and so one we found.

Best
Frank

[Dann Corbit]

Hi Dann,

that's better hardware I am using.

I made a test before I select the time for using!
For the first 1.000 of 26.146 positions (without upd.1).

30 seconds per move
60 seconds per move
180 seconds per move

Two questions:

1. How many other best / first moves engines found.
2. Different in evals (differents are very small, not interesting).

From 30 to 60 seconds I can see clearly better results, so I am using 60 seconds. At first, I am thinking 30 seconds are enough.

From 60 to 180 seconds lesser as 1% other moves engine Fire found. So, I am thinking time for analysis on my hardware with 60 seconds is good and I can start 10x 41.614 = 416.140 analysis.

Maybe in 3 years with 5x stronger Hardware and 100 Elo more for each of the engines we can produce a better result?!

Interesting is that 10 engine rejected in team work. With such a net you can be sure that the database is clear. Much more interesting as more time are many engines for Analysis. A Group of engines can found more as 2-3 engines with more time. That's the deal ... FEOBOS = teamwork by 10 engines!

Note:
6 Cores 10 Threads, 4.3Ghz with 60 seconds ...
No rating list will have such conditions in the next 3 years. Analysis will have the Quality important for rating list producer.

For us is the current result not more as a current rating list. Much more important is the Excel development because this work can be used again and again.

The Excel file we produced, with more as 11 million of Formulas, have very new and innovative idea. That was the most and a very hard work.

Between Klaus Wlotzka and myself 1.700 emails switched. Often, we solved problems 03:00 in the night. Most work we had with the ranking system.

A very hard work we do here in the last year. Much more work as in times I switched such development Information with Martin Blume for Arena.

Best
Frank

But your Hardware is great ...
You can see in KECR Excel that Houdini and Andscacs are most interesting for opening analysis. We can learn a lot to engine if we put over 41.146 positions a template. So many unknown details to engines, ECO codes and so one we found.

Best
Frank

According to my experience, 36 plies is bare minimum for opening book analysis.
And if the position is a gambit, really at least 40 plies are needed.

The issue is that you can find stunning upsets. I gave one from your book a little while back:
http://www.talkchess.com/forum/viewtopi ... highlight=

The position looks fairly reasonable until depth 29 when suddenly it jumps from +0.7 pawns to +2 pawns advantage.

If you include a position like that and it gets used it will likely give a win to a much weaker engine and bend the numbers a little bit.

Personally, I have come to the opinion that positions with an eval of zero after a deep search are also bad. These will end mostly in draws unless we can see evidence that it is really just a highly balanced position with an equal number of wins and losses.

These "hardwired draws" will also slant the results because they will cause draws for a weaker engine.

[Frank Quisinsky]

Hi Dann,

same opinion to draws and around 30% of the gambits.

We are working very tricky to produce with FEOBOS a database free of positions, produced fast draw games.

1. To located bad moves is easy.
2. To located drawish positions very complicated for us with Excel. But we found a way with Contempt and the ranking system. For the database with move transpositions we used a second way to located drawish positions (not the ranking system).

I know that you collected games and I know that you have very big databases with great material. I got from yourself for many years a B01 database, a nice gift for myself, thanks again!!. I hold it all the years in one of my main directorys and do a lot with it, because many games I missed in my own collected databases I found in yours.

I am thinking that 10% of Gambits today are interesting only, maybe 25- 30% are playable. We have around 30 ECO Codes ... three moves after ECO code was formed isn't enough ... same for Gambits. So we need more hardware power more plys. Thats right, but again I can't see that engines found not many better moves if I looking in different 60 seconds to 180 seconds with 10 threads and 6 cores.

For engine-engine all this isn't very intereting. Only ... engines can start with a balanced position, fast draw not possible, bad moves not included.

Today the interest on Gambits isn't very high. For most good known Gambits we found all what we need in strong GM analysis. Most Gambits are bad if the opponent know the idea. Today most engines have no problems with such Gambits and can calculate the right answere very fast. Only in 10% I can see that different engines have problems ... think so!

Looking in the past ... around the year 2000. We found out that most programs have Problems with dutch systms, french systems. Today dutch systems are for engine like Houdini, Komodo, Stockfish not a Problem. Stronger engines will not lost fast in french systems. And the same we can see for a lot of Gambits.

After my stats engines rejected 71% of gambits before and a lot of Gambits in the 41.614 database later after 10 engine analysis.

Again, FEOBOS can be better with faster Hardware and better engines in 3 years. Excel development is ready and to added new results easy.

But never I saw that what we do with FEOBOS before. I can't found over 500.000 engine Analysis for an opening development on other sites.

In my opinion, no better opening book for eng-eng can be produced as 10 engines do that in teamwork for us testers of engines.

Best
Frank

[Frank Quisinsky]

Other example:

user have a machine with 20 cores.
Now user is thinking ... best engines and all cores for the analysis. User is thinking, I have an Advantage in correspondence chess to the others with faster hardware.

WRONG ... if many pieces on board.
RIGHT ... if lesser pieces on board.

I can produce with 10 different engines and one core and the half of time a much better result in Teamwork of engines as the user with one engine and 20 cores with double time in most of cases (with many pieces on board).

So engine with 20 cores and ply 40 can't more see as 10 different engines with one core and maybe ply 32. Because engine with 20 cores and ply 40 can often not find a the right way ... for an example ... in ply 26.

From ply 32-40 engine analysed material not important for the right move in ply 26. So you need many different engines for a stronger result with very own ideas in programming.

My personal opinion after all what I saw in most of cases if ... you have many pieces on board.

If you have lesser pieces on board, situation is an other. More plys better results in most of cases.

For opening analysis absolutely clear for myself ...
We need many engines for optimate opening books and such things we developed in Excel. Stats on FEOBOS Start worksheet (new rejected lines by engine) is logical.

Sure, that with a better ranking system Klaus and myself developed the results can be better. Maybe a group of programmers can do it better as we try at first.

The stronger the engines the more intersting are analysis of opening systems. Can be a main topic in the future for computer chess. In the past often boring for the most of computer chess people.

Best
Frank

With other words:
2 engines with 10x better hardware I am using for FEOBOS and 4x more time can't produced a better results as 10 engines can do in Teamwork with the Hardware I am using. If the playing style of engine is different!! Speaking from many pieces on board, opening analysis. This is one of my personally FEOBOS results, means my opinion today. I learned here a lot in the last year.

[pohl4711]

I'd like to use the new SALC books as well; however, the larger of the two consists entirely of lines that at 10 moves are a bit too long for rating lists. The general preference these days seems to be for much shorter books, like Noomen's 2-moves and 3-moves. I'll probably start using the smaller closed-positions-only test suite; although quite small at 500 lines, the premise appears very interesting.

For ratinglists, 500 lines of the small SLAC V5 full-closed set should be enough (1000 games in an engine head-to-head are possible). And these lines are only 8 moves deep, not 10, like the big SALC V5 book.
The problem is, that SALC positions are very rare. I filtered the 25000 SALC V5 positions out of the BigDatabase 2018, which contains more than 7.2 million games(!!). Mention, that the new developed additional half-closed (full-closed) filters are lowering the number of matches, too. And that Komodo filtered out very bad positions (eval in the interval of [-0.6,+0.6]). And that all double-lines (doubles in the endpositions) were deleted. I had to fill up the SALC V5 positions with some 0.61-evals, in order to reach at least 25000 lines (out of 7.2 million human games)...

And 10 moves are not that much: From the startposition, play 2 pawn moves (centerpawns forward), 4 moves to bring 2 knights and 2 bishops in the game and the castle-move (which must be played in a SALC-position, of course), then you get 7 moves already...

I can only say: Give SALC a try - you will see much more thrilling computerchess, than you have ever seen before! Not only a much lower number of draws...

[Graham Banks]

ChessGUI version:
http://kirill-kryukov.com/chess/discuss ... p?id=41306

[Frank Quisinsky]

Hi Graham,

good, thank you!
And for yourself .. have a nice weekend!

Best
Frank

[Graham Banks]

.. have a nice weekend!

Best
Frank

Thanks Frank. Likewise to you.