Fruit reloaded seems to benefit from stockfish null move

[mar]

C'mon Dann

[PK]

Look at average opponent's ratings in Rybka forum post. Something seems not quite right there, as it is much lower than in case of the old versions.

Having said that, recently I tested the same formula within Rodent - it helped, but of course not in such a spectacular manner.

[elpapa]

This is what I get after 62 games at 90"/0.5".

Code: Select all

1 Fruit Reloaded            :   17   62 (+ 21,= 26,- 15), 54.8 %
2 Fruit Reloaded - Dann     :  -17   62 (+ 15,= 26,- 21), 45.2 %

[LucenaTheLucid]

Wow Dann,

If this is the case with Fruit, I wonder what would happen with Senpai? Or does it already use this type of null move pruning?

[Joerg Oster]

This is what I get after 62 games at 90"/0.5".

Code: Select all

1 Fruit Reloaded            :   17   62 (+ 21,= 26,- 15), 54.8 %
2 Fruit Reloaded - Dann     :  -17   62 (+ 15,= 26,- 21), 45.2 %

Yes, same here.
No improvement after about 250 games, test still running.

[syzygy]

I agree with Kai that if such a small change was tested to give a 500 Elo improvement, something went wrong while testing. Since the test result is bogus, there is no reason to expect the change to be an improvement.

[Isaac]

The given results in the 1st post make no sense to me.
The engine won all its game, 0 draw. Its elo should be infinite right? Not 3300. And what's with this 0 error bar?

[Joerg Oster]

Indeed.
500 games at 20"+0.08", 64 MB Hash each:

Code: Select all

Finished game 499 (Fruit-Reloaded-DC vs Fruit-Reloaded): 1/2-1/2 {Draw by 3-fold repetition}
Score of Fruit-Reloaded-DC vs Fruit-Reloaded: 113 - 143 - 244  [0.470] 500
ELO difference: -21
Finished match

Fruit-Reloaded-DC = Fruit-Reloded with SF Null formula

Not very convincing!

[mcostalba]

C'mon Dann

Yeah, Dann is famous to report grossly inflated results.

I remember few years ago it did it with some kind of formula, and once the retest was clearly showing a much less specatular result he said: "So? I have raised tha attention bar and made people to retes it".

...so nothing new here

[PK]

If I understand Fruit Reloaded code correctly, it has been built around exactly opposite principles than Stockfish. Reported depths are low, but search is extremely reliable and finds some tactical shots surprisingly fast. Most of the prunings/reductions are switched off or modified on moves with high history counters. Unification of late move reduction and late move pruning under the same hood is a thing of beauty (I hope Daniel Mehrmann is reading this!) but it makes tuning of this feature extremely hard. Inserting the elements of fast paced Stockfish search into this slow but sure framework is unlikely to succeed.