Updates Cerebellum

[pohl4711]

https://zipproth.de/Brainfish/download

Both Cerebellums updated. Lc0-Cerebellum was recalculated with TCEC Superfinal winner Net (T40.T8.610) and is +40% bigger, than before.

I started the testrun with BrainFish (Stockfish-patch date 190531, so the result is comparable with my latest Stockfish-testrun). Result not before next Monday.

https://www.sp-cc.de

Stay tuned!

[giovanni]

https://zipproth.de/Brainfish/download

Both Cerebellums updated. Lc0-Cerebellum was recalculated with TCEC Superfinal winner Net (T40.T8.610) and is +40% bigger, than before.

I started the testrun with BrainFish (Stockfish-patch date 190531, so the result is comparable with my latest Stockfish-testrun). Result not before next Monday.

https://www.sp-cc.de

Stay tuned!

I understand that Cerebellum is built only upon engine evaluations. It is a balanced graph of chess positions with consistant scores of all nodes, i.e., the evaluation is backpropagated to the root.
In this regard, how Cerebellum differs from a well mini-maximized opening book?

[pohl4711]

In this way: “Furthermore the Cerebellum Library is able to handle position repetitions by with re-visiting some nodes and then playing different moves.“
Avoiding those possible move-loops is the difference.

[giovanni]

In this way: “Furthermore the Cerebellum Library is able to handle position repetitions by with re-visiting some nodes and then playing different moves.“
Avoiding those possible move-loops is the difference.

Thank you. But I am a little confused on how is this related to the incredible strength of the Cerebellum approach, that made it a real breaktrough in opening book building.