Unbalanced Human Openings released

[MMarco]

Thank you for your contribution to the community. I'm eager to try your new set, when test engine I watch quite a bit of games, so new interesting positions are always welcomed. Very nice.

[jefk]

reply to mr Pohl,

no misunderstanding, i thought it was mostly for comp-games
so maybe i was talking in a bit ironic way (putting
emphasis on human chess, but i know compchess
also is a serious 'discipline' (also used btw in corresp chess )
and indeed then as an area to seek further improvemnt(s)
and chances to win you need unbalanced openings
to avoid dead draws, fully agree.

As i wrote earlier, being a lazy human chess player, but
also knowing about computer chess, i sometimes try to
learn from computer chess (and projects like yours)
to improve my own game(s).

[MikeB]

I released my Unbalanced Human Openings.

Unbalanced Human Openings contain:

- 100% moves played by humans, only. Both players had at least 2300 Elo. Out of the Megabase 2020
- no manually constructed openings (like my Drawkiller openings)
- no manually added moves to make castling impossible (like my NBSC openings)
- no selection of piece-patterns (like my SALC openings (white and black castle to opposite sides of the board))
- no double endpositions
- all endpositions analyzed by Komodo and filtered by that Komodo eval-value.

Learn more on my website:
https://www.sp-cc.de/unbalanced-human-openings.htm

Direct download:
https://www.sp-cc.de/files/unbalanced_h ... s_v1.0.zip

Enjoy!

+1 Thank you - I really appreciate what you do and that you freely share it.

[pohl4711]

I like this very much, congrats! I've used my own version of this idea for quick tests for a while now, but only with a small number of openings. I'm a bit surprised that your score range is so high (0.80 to 0.89 Komodo eval), as I've generally figured that the win/draw line was around 0.70, but your results show about 50% draws so it must be about right. Perhaps the explanation is that since these are openings played by strong human players they are probably not as bad for Black on average as the evals say.

If anybody is interested in my pre-testing results, which I did, to find out, which Komodo-eval works best, here the results. The 6 moves deep raw-data set was used (and filtered by the Komodo-Eval interval)

Each testrun, as usual:
2'+1'', Singlethread, i7-8750H Hexacore mobile CPU, 256MB Hash, cutechess-cli (no TB for engines, but 5 Syzygy for cutechess), Contempt=0 for all Stockfish. All openings replayed with reversed colors. Round Robin with 1500 games with SF 11, SF 10, SF 9 and SF 8. Each SF played 250 games vs. each of the 3 opponents = 1500 games per testrun.

Komodo Eval [+0.50;+0.59]: Draw-rate: 62.7%, Elo-spreading: 152 Elo (Advanced Armageddon scoring: 252 Elo)
Komodo Eval [+0.60;+0.69]: Draw-rate: 57.6%, Elo-spreading: 157 Elo (Advanced Armageddon scoring: 275 Elo)
Komodo Eval [+0.70;+0.79]: Draw-rate: 54.7%, Elo-spreading: 160 Elo (Advanced Armageddon scoring: 289 Elo)
Komodo Eval [+0.80;+0.89]: Draw-rate: 49.7%, Elo-spreading: 163 Elo (Advanced Armageddon scoring: 308 Elo)
Komodo Eval [+0.90;+0.99]: Draw-rate: 47.9%, Elo-spreading: 155 Elo (Advanced Armageddon scoring: 283 Elo)
Komodo Eval [+1.00;+1.09]: Draw-rate: 41.5%, Elo-spreading: 147 Elo (Advanced Armageddon scoring: 280 Elo)

As you can see, the "peak" is [+0.80;+0.89]. Best Elo-spreading here (with both classical scoring and Advanced Armageddon scoring). With higher evals, the draw-rate decreases (of course) more and more, but with higher evals, you get more 1:1 pairs, too (an opening is won for white both times, when Engine A and B play it (and repeat the opening with reversed colors)). And 1:1 pairs are as bad for the Elo-spreading, as draws are (in classical scoring). Thats because the Elo-spreading in such a RoundRobin-tournament is the best parameter for measuring the quality...

These testruns were the reason for me to choose the Komodo Eval [+0.80;+0.89] for my Unbalanced Human Openings (and the Komodo Eval [+0.70;+0.99] for the wider and bigger 8moves-openings set, which contain more than 30000 lines).

[pohl4711]

I have just downloaded the brand new and much stronger (+200 Elo?) Komodo Dragon engine, which has an nnue-net, now.
I will do a complete new evaluation of my Unbalanced Openings raw-data sets and will build new Unbalanced Human Openings V2.00 out of this new data.
But the 4,6 and 8 moves raw data sets contain 386848 opening-lines. The evaluation with 15 seconds per endposition will take more than 2 months (!). And after that, a lot testruns are needed. So the release will not be before February 2021. If all works as expected.

Stay tuned!