AB/NN Engine?

[Eduard]

Would not be more productive and better, to construct an AB engine that use a trained MCTS neural network (maybe like a Leelafish), but only from move ~16 to ~35?

For the first 15 moves I favore using a own opening book, and after move 35 I favore an AB search with endgame tablebases. For the middle game I could use different neural networks, according to the style (!) of the opening book used. Would not such an engine be fantastic?

[peter]

Hi Eduard!

Would not be more productive and better, to construct an AB engine that use a trained MCTS neural network (maybe like a Leelafish), but only from move ~16 to ~35?

For the first 15 moves I favore using a own opening book, and after move 35 I favore an AB search with endgame tablebases. For the middle game I could use different neural networks, according to the style (!) of the opening book used. Would not such an engine be fantastic?

Good idea too to me.

Another one: how about letting LC0 or an engine like this run on GPU together with SF or an engine like this on CPU and give the A-B-engine order by GUI or UCI- option of A-B-engine to add a second primary variant in MV- mode after LC0's choice, in case of disaggreement between the two engines?
Decision about move to choose comes from A-B-engine or from NN- engine as one of two possible settings, if LC0 being ordered to ponder about A-B-engine's different move, but I doubt NN-engines' ability of hash-learning compared to SF's one.

Of course in analysis mode I'd still prefer having both engines run by themselves and try Forward- Backward of moves of disagreement. Nucleus- like constructs could serve as game- playing engines too. Anybody tried this one with late LC0 and SF as slave and master?

[Eduard]

Dear Peter, your idea is good too. I know you are a correspondence chess player. In any case, it makes sense to use the CPU and the GPU in parallel. It gets the maximum out of the hardware! I love playing on playchess (with own opening books). An engine which got the maximum out of the hardware, is of course the best! The originality of an engine does not interest me.

[Steppenwolf]

I also think this is a good idea to go foreward in the steps of LeelaFish.
Why not combine the best from both engines in one code: AB/Stockfish/CPU and NN/Lc0/GPU in one running simultansously?

Programmers requested!

[velmarin]

You can try ChessCombi, and see if the idea works

In readme:

**************If EnableSmartSwtiching is 0:****************
-> NumberOfEngine1Moves=VALUE
-> NumberOfEngine2Moves=VALUE
*VALUE is a number.
*The NumberOfEngine1Moves dictates how many moves Engine 1 can play before swtiching.
*The NumberOfEngine2Moves dictates how many moves Engine 2 can play before swtiching.
*Example: If NumberOfEngine1Moves is 10, and NumberOfEngine2Moves is 5.
*Engine1 plays for 10 moves, then switches to Engine2. Engine2 plays for 5 moves, then switches back to Engine1.

****************************END****************************

[Eduard]

Good idea, thank you! Runs this under a Fritz GUI?

I found this two links

https://chesscombiengine.blogspot.com/2 ... s.html?m=1

https://chesscombiengine.blogspot.com/2 ... t.html?m=1

I like to test it. Where can be downloaded a current version v4?

[Eduard]

I downloaded V1 and tested it, it works fine under Fritz 15. Now I only need a good GPU. Then I can combine Stockfish with Lc0 as I like it. And with full power. Cool thing, keep it up.

[Eduard]

Here is a link to a previous posting with download link. Sorry, I only found the message now.
viewtopic.php?t=69135

[Steppenwolf]

Great Idea.
Is it possible to make this also available for macOS Mojave?

[Eduard]

A suggestion to the developer of ChessCombi: It would be good to see (maybe in the display?) when calculating engine A or engine B.