do modern programs play 22.Nxf7 in this position...

[pilgrimdan]

Andscacs 0.90 after 8 seconds:

Code: Select all

info depth 22 currmove g5f7 currmovenumber 14
info depth 22 seldepth 35 score cp 95 lowerbound nodes 7822002 nps 949502 tbhits 0 time 8238 pv g5f7

And my current dev, with threat evaluation rewritten and retuned, 4 seconds:

Code: Select all

Andscacs 0.90087
info depth 21 currmove g5f7 currmovenumber 37
info depth 21 seldepth 35 score cp 104 lowerbound nodes 3796740 nps 934927 tbhits 0 time 4061 pv g5f7

looks like you have done really well with Andscacs...

[cdani]

And my current dev, with threat evaluation rewritten and retuned, 4 seconds:

looks like you have done really well with Andscacs...

Threat stuff on Andscacs was quite primitive. It needed a good rewrite. It helps a lot to find faster tactical stuff like this one.

[brtzsnr]

Zurichess finds Nxf7 relatively fast and it's weaker than top engines:

info depth 21 seldepth 26 multipv 1 score cp 169 nodes 33445791 time 15288 nps 2187601 pv g5f7 f5f7 e3f3 e7f6 h5h6 f7e7 e4f6 g7f6 e2e4 e7d7 f3f6 c6a5 e4g4 g8h8 f6e6 e8e6 e1e6 a5c4 g4e4 d7d8
e6e7 d8g8 e7h7 g8h7 e4a8

[mar]

well, Cheng is not top at all but finds Nxf7 in less than a second

[cdani]

Well, Cheng is not top at all but finds Nxf7 in less than a second

I hope you try to improve it at some point! There are some known techniques that will add 50-100 elo to it quite fast.

[mar]

I hope you try to improve it at some point! There are some known techniques that will add 50-100 elo to it quite fast.

Thanks Daniel
I'm aware of only two techniques that add a lot of elo, one is nullmove and another lmr. I also got 100 elo thanks to Texel tuning evaluation.
Everything else was about accumulating enough small improvements.
So I wonder which known techniques you mean? Perhaps I'm missing something?

[cdani]

I hope you try to improve it at some point! There are some known techniques that will add 50-100 elo to it quite fast.

Thanks Daniel
I'm aware of only two techniques that add a lot of elo, one is nullmove and another lmr. I also got 100 elo thanks to Texel tuning evaluation.
Everything else was about accumulating enough small improvements.
So I wonder which known techniques you mean? Perhaps I'm missing something?

Yes, I mean also accumulating some things that should work on most engines. For example I compared it to Andscacs and I can suggest at first sight:

* Reduction of null move can be quite more ellaborated, and also can call directly to quiescence.
* Eval pruning - child node.
* Futility pruning - child node.
* Singular extension.
* In your current futility pruning your should be able to increase a lot the depth at which is applied.
* See pruning
* Different flavors of refutations.
* Update the history on tt cutoff.

I suppose that if you can apply those it will win a lot. I know is not that easy

[mar]

Thanks for the tips, I'll definitely try some of them.

[PK]

Rodent gets it in one second. Perhaps this is one of those deceptive positions where engines with better pruning/reductions perform worse.

[petero2]

I hope you try to improve it at some point! :-) There are some known techniques that will add 50-100 elo to it quite fast.

Thanks Daniel :)
I'm aware of only two techniques that add a lot of elo, one is nullmove and another lmr. I also got 100 elo thanks to Texel tuning evaluation.
Everything else was about accumulating enough small improvements.
So I wonder which known techniques you mean? Perhaps I'm missing something? :)

Yes, I mean also accumulating some things that should work on most engines. For example I compared it to Andscacs and I can suggest at first sight:

* Reduction of null move can be quite more ellaborated, and also can call directly to quiescence.
* Eval pruning - child node.
* Futility pruning - child node.
* Singular extension.
* In your current futility pruning your should be able to increase a lot the depth at which is applied.
* See pruning
* Different flavors of refutations.
* Update the history on tt cutoff.

I suppose that if you can apply those it will win a lot. I know is not that easy :-)

In that list my guess is that singular extensions gives the largest gain. In texel SE was worth around 25 elo. This assumes you use the modern version of singular extensions though (called restricted SE in the CPW), not the historic ChipTest/Deep Thought version which seems to have a too large overhead to be effective.