FCP Tourney-2020: 41.000 games 20+5 to play on Intel i9-10900k

[Frank Quisinsky]

Hi,

no problem Jon!
Again, each position are analyzed by 10 engines!
So I have to each FEOBOS position ranking and statistics.

I will search the position ...
Need a while, must make me a strong coffee before ...

Frank have an order ... juchu!
I like that!



A bit later this evening, after the coffee ...

Best
Frank

[Frank Quisinsky]

Oh, I overwrite my own message to Arasan!

Arasan is for myself the biggest competitor for Wasp.
The day will come, Wasp is stronger.
The 80 Elo advantage from Arasan to Wasp are to search in the better endgame Arasan have.

The problem is:
Arasan today isn't only strong in endgames. The style from the last versions are more and more interesting. Arasan today can play all the playing phases very equal and is today much more danger with many pieces on board. Much more stronger in aggressivity. The playing style from Arasan today is very nice and interesting.

Not a problem, a problem for Wasp.

Arasan today is a wonderful engine (I know that) but in my brain more or less a hate-love.
Hate-love because I like to see that Wasp is stronger as Arasan.

Not the opinion by John Stanback, my own opinion!
Not easy for Wasp to beat the Elo strength from Arasan.

Loud thinking ...
Now the coffee and than I will search the FEOBOS position!

[Frank Quisinsky]

Position in database (from 41.614 FEOBOS positions) = 17.668
Contempt 3 used the first 33.009 rankings ...

Ranking place = 9.953 (ranking system for the 41.614 positions).
Opening = D20
68 D20 lines in FEOBOS database
from D20 openings = Place 21 of 68

And here the analyzes (I have from all 41.614 FEOBOS postions - x10 engines - the output analyzes files, I can give it to if you like).
I have to search for position 17.668 only ... but you can download the complete FEOBOS material.

FEOBOS = 3 moves after ECO-Codes formed.
D20 formed after three moves, so the FEOBOS lines to D20 goes 6 moves deep.



Give me an other position ...
Again, the opening book is bugfree!!

[Frank Quisinsky]

Short hint:
Developers from Microsoft gave the information:
Not more as 5 Millions of formulas possible in Excel or Excel produced problems.
Klaus Wlotzka mananged 16 Millions of formulars with a lot of Makros.

Such things and many other information about each of the positions, ECO codes produced FEOBOS.
Project ended 2018 ...

But all the material is available on my website.
http://www.amateurschach.de/main/_new-opening-book.htm

Best
Frank

PS: Looking the nice chessboard ... created with Excel!

[Frank Quisinsky]

To Rybka ...
Short game example with Rybka are bad.
Only 4-5 engines in TOP-41 produced more fast loses as Rybka.
From TOP-Engines Rybka ist on the last place in king safty, understand in many opening systems nothing.

Rybka is only good in transposition into endgame ...
Middlegame stats are to 99% the same Fruit produced.

Best
Frank

[Frank Quisinsky]

I forgot, of course we produced 2016 / 2017 FEOBOS with 10 engines.
Sure, the 10 engines with strongest opening stats are used!

Have a look in the list of engines!

[Frank Quisinsky]

I will give different examples for FEOBOS book lines:

Second best position (in ranking on place 2).



Ten engines gave not a 0,00 eval.
Ten engines gave not a bad eval.
10 engines found 8 different best moves with very balanced evals.

SENSATIONAL, such a test postion for engine-engine matches!!
A big tree can be produce with such a great balanced position (good for both sites).
ON TOP-2 from all opening lines we have (3 moves after ECO codes formed).

Next example:



Ranking 40.000 of 41.614 ...
Three of ten engines gave bad evals.
Positions not good for engine testing because one of the engines started with a to high advantage.

Next eample:



Bad for engine testing, because three engines gave 0,00 eval.
Probability for a fast 3-fold draw is to high!!!
With higher Contempt the FEOBOS book avoid draw.

With Contempt = 3 = 0,35% of engine-engine games ended with a draw below 20 moves.
With Contempt = 4 = 0,31% of engine-engine games ended with a draw below 20 moves.
With Contempt = 5 = 0,28% of engine-engine games ended with a draw below 20 moves.

For Contempt 2 = 0,48% (to high).

Optimal for engine testing is Contempt 3 with 33.009 positions, enough balanced material to all of the 500 ECO codes are inside, 3 moves after ECO code formed. This book I am using for FCP Tourney-2020 but with grandmaster stats to popular openings. So the FEOBOS 20.1 Contempt 3 book try to produced / simulate most popular opening systems the current 860 best players of the world like to play. I think I should try to produce material with engine-engine matches interesting for grandmasters. And I know three grandmasters follow the tourney for the material FEOBOS / FCP Tourney-2020 produced.



The last position is interesting.
With Contempt 2 the position is in FEOBOS database for create opening books.
Position can be on the chessboard if FEOBOS 20.1 Contempt = 2 opening databases or books will be used.
Books to different books format and all *.epd, *.pgn files (sorted with ranking system), developments files, German docu ... all what we produced can be download on my site.

Best
Frank

[Frank Quisinsky]

Hi there,

now the results from Round 18 of 50 are online!

Frank's Chess Page
https://www.amateurschach.de

Best
Frank

[Frank Quisinsky]

Hi there,

on https://www.amateurschach.de I have added first grafics from the work Klaus Wlotzka do here.
Klaus Wlotzka is an expert in Excel and with Excel we will find out more about playing styles and other interesting questions around the FCP Toruney-2020. Most of work we will do after the tourney. So we have more material for better statistics.

During the tourney, easy statistics will be available only!

But you can see that will be a very powerfull tool what we have in our brain to do here.

Best
Frank

Tomorrow round 19 will be ended.
First results (tournament table from Round 19) I added on the detail page.

[jonkr]

I will add a UCI contempt option for next version, I think it would be fairly straight-forward. I've been slow to add contempt because I think most people like to play with the strongest engines, so in any crowd that includes Stockfish, Leela or any variations on those engines, it would give worse results. It also isn't always the most pretty, I notice games where Stockfish sorta shuffles pieces around for a while looking for a weakness until eventually the opponent goes wrong, but it is better than quick draw by rep, and for Stockfish at least quite effective.

What's also kinda interesting is that draw rates in self-play for Slow Chess have gotten super high so maybe contempt would actually help testing, not sure. But against a different engine draw rates suddenly become way lower, like in an 8000 bullet game FRC gauntlet the SlowDev version got 60.2% draws versus Slow 2.2, and only 42.8% draws versus Stockfish 7 despite almost the same win %.

I do find the variety of engines interesting, although at current point seeing where games go wrong is almost always beyond my abilities. Back when Senpai 2.0, Wasp, and SlowChess were similar rating I did notice that Senpai seemed most vulnerable to getting its king attacked. (I was still hand-tuning back then and learned from watching Senpai's endgame I had underestimated the scoring advanced passed pawns and opponent king distance, and overestimated passed pawns still on ranks 2-4. Separately I also realized that it wasn't bad to have King Safety values that could be wrong from being too high, which I had initially avoided. Instead it was better to try to see danger early and estimate, and find more attacks. Also I first read about the "Space" evaluation for advancing pawns in opening to leave space for pieces behind them in Senpai 2.0 update notes.) I do kinda miss that level, since there was more variety of opponents to choose from, and easier to progress from just trying out ideas eval or search from looking at games or test positions. (Update for those curious : my current dev version scores about 10-15 elo above 2.2, so not sure what to my plans for releasing are, as that isn't enough to say for sure if it would even score higher on all lists.)