SPCC: Testruns of BrainFish 160610 and 160712 finished

[pohl4711]

Testruns of BrainFish 160610 and BrainFish 160712 finished.

Endless RoundRobin tournament with long thinking-time games of BrainFish 160712 updated, too...


http://spcc.beepworld.de

(Perhaps you have to clear your browsercache or reload the website)

[Vinvin]

What's BrainFish ?

[Leto]

BrainFish is Stockfish with an internal opening book which is a light version of the cerebellum opening book, coded into the engine. The engine is just one exe.

In analysis mode the internal opening book is disabled, so if you want to analyse the opening phase you can.

You can still use your own opening book in tournaments, BrainFish will follow whatever opening book you set for the tournament and will only start using its internal book after the last move of the tournament book is played.

Here is the website:
http://www.zipproth.de/

[pohl4711]

BrainFish is Stockfish with an internal opening book which is a light version of the cerebellum opening book.

The Cerebellum Library is much more, than an opening book! See the diagram below, to understand, how its backward-calculation/evaluation works.

[Vinvin]

BrainFish is Stockfish with an internal opening book which is a light version of the cerebellum opening book.

The Cerebellum Library is much more, than an opening book! See the diagram below, to understand, how its backward-calculation/evaluation works.

http://spcc.beepworld.de/files/library_erlaeterung.jpg

What's the size of the engine executable ?
Is it only stronger in theoretical opening ?
Same strength as the regular SF in chess960 ?

[pohl4711]

BrainFish is Stockfish with an internal opening book which is a light version of the cerebellum opening book.

The Cerebellum Library is much more, than an opening book! See the diagram below, to understand, how its backward-calculation/evaluation works.

http://spcc.beepworld.de/files/library_erlaeterung.jpg

What's the size of the engine executable ?
Is it only stronger in theoretical opening ?
Same strength as the regular SF in chess960 ?

File-Size is around 35 MByte (4.4 million positions in the Cerebellum-Library)
And Brainfish is only stronger in regular chess with normal openings. In chess960 the Cerebellum-Library has no effect.

[Ozymandias]

Brainfish is only stronger in regular chess with normal openings.

The internal book is more like an iDea tree, than an opening book, then?

[Thomas Zipproth]

The internal book is more like an iDea tree, than an opening book, then?

Yes, that's true. It has some improvements in calculating the single positions and in transposition handling, but I don't know iDea in full detail, so I can't really compare then. But the basic Idea is a an opening book created from the engine itsself, which is a spinoff from the Cerebellum Library created for the deep Analysis of Chess positions and opening lines.

[Ozymandias]

If Cerebellum evolves to accept Aquarium epd results, works with any engine under the InfinityChess GUI and offers flexibility, when compiling the book, this could be interesting.
iDea is geared towards analysis, not engine gameplay, so there's a window of opportunity there, for a complementary tool.

[Lyudmil Tsvetkov]

Without wanting to underestimate Thomas' effort in any way, this is more a book than an engine. No brain behind all in-built innovations, just meticulously and sequentially evaluated game tree.

All elo increase is due to:
1) saving ( a lot) of time on book moves, especially considering the length of the book, and even more the very short TC.
2) potentially picking up stronger lines, bearing in mind the backward-investigation approach, moving closer the game horizon, and thus facilitating perfectionised decision-taking.

At 5' + 3'' you would witness halving of the elo gain, at LTC elo gain migth be minimal.

An even better engine would be an engine with in-built all precalculated 32-men TBs.

Also, the BF approach contradicts the urge of most modern top engines to discard books/longer books and concentrate on improving early opening gameplay.

What I would grant BF though, is that the backward-generated book possibly indeed picks better moves in a range of situations than standard SF, but in no way best possible moves. For example, standard SF rightly picks 1.e4 as best, while BF goes wrong with 1.d4, for whatever reason.

I could test ASM and BF out of curiosity, but would in no way place them in the same standings, as playing conditions are simply unequal.

Again, this is just my personal opinion, I appreciate all efforts, and would not like to belittle Thomas' work in any way, nice to have new tries, but in this case the playing field is simply unequal for a standardised rating list.