Introducing the *.EBF project

[Henk]

Skipper currently only counts material.

[But actually it should only count if King is captured for the deeper the search the less evaluation features are needed and evaluation only costs time]

[Dann Corbit]

Suggestion:
Depth searched for the nodes should be a function of how frequently the node is played.

For instance, if the node appears 1000 times in real games, then 40 should be the minimum {goal} depth.

If a node appears 100 times, then 35 should be the minimum depth.

If a node appears 10 times then 26 should be the minimum depth.

Those are goals, of course.
And a stronger search by engine X for position Y should simply overwrite its previous entry.

After the book is filled out, it should be mini-maxed. It would also be good if the mini-max engine was aware of a need to search a missing node with a high play rate or a high score and a low depth and write out a request for a deeper search to a log file.

[Rebel]

These are nice ideas to experiment with.

Another one is -

I am currently analysing the 6.5 million selected *.EBF positions with SF7 at 2 seconds. When that is finished I am considering the idea to load them into the TT at program start and use them during search. Of course ProDeo evaluation will collide with the SF7 evaluation and I expect a regression, but I want to know anyway.

A better approach (I think) is to analyze the 6.5 million positions with ProDeo itself and use it during search.

[Dann Corbit]

These are nice ideas to experiment with.

Another one is -

I am currently analysing the 6.5 million selected *.EBF positions with SF7 at 2 seconds. When that is finished I am considering the idea to load them into the TT at program start and use them during search. Of course ProDeo evaluation will collide with the SF7 evaluation and I expect a regression, but I want to know anyway.

A better approach (I think) is to analyze the 6.5 million positions with ProDeo itself and use it during search.

This idea is very much like my permanent table of pv nodes.
After all, moves which were actually played are pv nodes.
They all have exact scores and are very important in the tree.

[Rebel]

Yep.

There is ELO in big data extracted from comp games.

[Dann Corbit]

I also analyze the games of every big tournament as they happen.
I have the first 3 rounds of Tata Steel tournament analyzed, for instance.

The reason I bother with that is that the players in these tournaments will be right on top of the latest theory and will have prepared very diligently. So we can find important novelties and creative positions.

[Rebel]

First *.EBF results are in.

And are good.

http://rebel13.nl/prodeo.html#results

Not sure what to expect. The diminishing return effect will cause smaller ELO gains, that is a given, you just can't catch every reasonable playable position for 10 moves (in test currently), let alone for moves 11-15, the end goal of this experiment.

Anyway, this is going to be a looooooong ride.

[Henk]

Don't know if you can trust engines about opening moves. For instance if I analyse 1. g3 after say forty seconds it gives a value of -0.2 and after five minutes Stockfish 6 gives a value of 0 .. 0.06 for that move on my computer.

So first it gives you the idea that g3 is not a good. And who wants to wait five minutes for one evaluation.

[Dann Corbit]

Don't know if you can trust engines about opening moves. For instance if I analyse 1. g3 after say forty seconds it gives a value of -0.2 and after five minutes Stockfish 6 gives a value of 0 .. 0.06 for that move on my computer.

So first it gives you the idea that g3 is not a good. And who wants to wait five minutes for one evaluation.

I generally go to 40+ plies on any opening move I plan to use.
36 is bare minimum for something within the first 20 plies.

Lesser depths are OK as you move farther out.

The opening position of chess is a closed position. Those positions are the hardest for chess engines.
As more moves are made, the game usually opens up.

[Rebel]

Don't know if you can trust engines about opening moves. For instance if I analyse 1. g3 after say forty seconds it gives a value of -0.2 and after five minutes Stockfish 6 gives a value of 0 .. 0.06 for that move on my computer.

So first it gives you the idea that g3 is not a good. And who wants to wait five minutes for one evaluation.

First of all *.EBF books are meant to consult after you are out of your own book. Secondly, if you would compile the example source code with an EPD of the start position you will notice SF7 | K7 and G3 will play 1.e4 and not 1.g3