OKE - Opening Knowledge Engines

[Guenther]

1. you need to start later in the opening

I think it is completely irrelevant where you start. A fully closed position is difficult, like locked positions, but some computers can find good answers.

It is totally relevant because with only 2! up to 4 moves you are left with much too much equal moves, which are just a matter of taste/style.
2-4 moves is the current discussion for OKE, it is incomprehensible for me how you can dispute about this.
The whole discussion lacks of logic as most threads here, thus I leave it here.

Most chess positions have 2 to 4 good choices.

...

You completely missunderstood. We are talking about move choices after 2-4 moves not about 2-4 move choices, sigh...

[Dann Corbit]

My programs emit this:
rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - bm c4=10; e4=9; d4=9; Nf3=6; e3=3; g3=3;

[Dann Corbit]

1. you need to start later in the opening

I think it is completely irrelevant where you start. A fully closed position is difficult, like locked positions, but some computers can find good answers.

It is totally relevant because with only 2! up to 4 moves you are left with much too much equal moves, which are just a matter of taste/style.
2-4 moves is the current discussion for OKE, it is incomprehensible for me how you can dispute about this.
The whole discussion lacks of logic as most threads here, thus I leave it here.

Most chess positions have 2 to 4 good choices.

...

You completely missunderstood. We are talking about move choices after 2-4 moves not about 2-4 move choices, sigh...

One of us has completely misunderstood.

[noobpwnftw]

I have built a database using a slightly modified SF to adjust MultiPV depending on shallow search results.
It now has about 2 billion unique positions, each analyzed with depth >=22 and MultiPV >=5, my experience is that eventually which engine or parameters you use becomes irrelevant - it converges to statistics sooner or later.
One thing important to avoid is "double" selection bias: human games or most game sets are not well distributed among all openings.

[Dann Corbit]

Only two positions in the EPD files actually have valid e.p. tags.
[d]rnbqkbnr/1p1ppppp/8/pPp5/8/8/P1PPPPPP/RNBQKBNR w KQkq c6
[d]rnbqkbnr/p2ppppp/8/1Pp5/8/8/PP1PPPPP/RNBQKBNR w KQkq c6

All the others are spurious. Because of this, there are more than 1380 duplicates in the file.
About 18% are duplicates.

[Dann Corbit]

This one is strange, it is a mate in 2:
[d]rnbqkbnr/pppp1ppp/8/4p3/8/5P1P/PPPPP1P1/RNBQKBNR b KQkq -
Qh4+ g3 Qxg3#

It's a variant on the fool's mate.

[Dann Corbit]

I have built a database using a slightly modified SF to adjust MultiPV depending on shallow search results.
It now has about 2 billion unique positions, each analyzed with depth >=22 and MultiPV >=5, my experience is that eventually which engine or parameters you use becomes irrelevant - it converges to statistics sooner or later.
One thing important to avoid is "double" selection bias: human games or most game sets are not well distributed among all openings.

With over a billion positions analyzed, you can backsolve.
The depth value at the root will be approximately 22 + distance to leaf, (unless the root depth is already greater than this).

[noobpwnftw]

The depth value at the root will be approximately 22 + (distance to leaf).

Indeed, but there is still a horizon problem, although the chances of getting something wrong diminishes when depth goes higher.
The search breadth should be considerably wider than vanilla search from root position due to MultiPV at each child.
I use averaging back-propagation of engine eval to refine statistics of each move.

[Dann Corbit]

The depth value at the root will be approximately 22 + (distance to leaf).

Indeed, but there is still a horizon problem, although the chances of getting something wrong diminishes when depth goes higher.
The search breadth should be considerably wider than vanilla search from root position due to MultiPV at each child.
I use averaging back-propagation of engine eval to refine statistics of each move.

Have you done this for European chess or only for Xianqui?
I guess it would make a book that would destroy the Cerebellum version.

[noobpwnftw]

The depth value at the root will be approximately 22 + (distance to leaf).

Indeed, but there is still a horizon problem, although the chances of getting something wrong diminishes when depth goes higher.
The search breadth should be considerably wider than vanilla search from root position due to MultiPV at each child.
I use averaging back-propagation of engine eval to refine statistics of each move.

Have you done this for European chess or only for Xianqui?
I guess it would make a book that would destroy the Cerebellum version.

I have done both, although I have yet to make an user interface for Chess - online probing via API is already available.