use multiple neural nets?

[Warren D Smith]

Suppose two nets are trained, A and B.

A is slow, but smart.  Because its extra slowness outweighs its extra smarts, net B yields

a stronger chessplayer with equal thinking-time limits.  But because it is smarter,

net A would be stronger with equal node-count (time ignored).

Simple idea:

use net A for nodes (game positions) early in each search, then switch to net B.

E.g. the first 1000 nodes use A, then switch to B for the next 9000 nodes, then

make a move (for a 10000 node search).

There are many possible flavors of this idea.   The goal is to

gain the benefits of A's extra smarts, where it pays off more, i.e. for nodes that

play a "more important" role in the search; but gain the benefits of

thinking faster about the "less important" nodes.  Best of both worlds.

I have no idea whether this idea can be made to yield increased playing strength.

But it sounds plausible that it might.   And the question might be easy to answer

by test.

[Warren D Smith]

And by the way, one thing I long urged the alphazero team to do (but they did not)

was essentially this.

They were training net to input position, output (1) evaluation [expected game result] and

(2) probability estimates for each candidate move (moves take you to daughter positions).

I suggested ALSO train net to output (3) |error| estimates for the evaluation, i.e.

truer-evaluation is eval backed up by a search; error is the truer-eval minus the

no-lookahead eval.

There would be many benefits of that, which I cannot go into here.

But to name just one -- for the current thread's suggestion

of multiple nets, the availability of error estimates would enable you to, for example, 

(a) use the fast dumb net

(b) if error estimate unusually large, then re-evaluate position using the slow smart net.

[Threshold for b needs be tuned to maximize experimental playing strength.]

Result is you get the benefits of high speed and high smarts, having cake & eating it too.

[Shah]

There is a similar idea that is being heavily discussed here, which is to switch to an ender network when piece count gets below some value (endgame stage)
Seems that leela main problem is in the endgame tactics.
She suffers no time issues, as the "smart and slow leela" as you say, usually gains a significant advantage in midgame, despite wasting significantly less time compared to her opponent.
So no problem here I suppose.
Endgame on the other hand is a true weakness.

[Warren D Smith]

On Tuesday, December 25, 2018 at 6:20:23 PM UTC-5, Shah wrote:

There is a similar idea that is being heavily discussed here, which is to switch to an ender network when piece count gets below some value (endgame stage)
Seems that leela main problem is in the endgame tactics.
She suffers no time issues, as the "smart and slow leela" as you say, usually gains a significant advantage in midgame, despite wasting significantly less time compared to her opponent.

--yeah, well, leela still appears weaker than stockfish.  What I am proposing could conceivably deliver similar strength to leela but

with a lot faster runtime... which might be enough to get past stockfish.  I mean, leela but 10X faster?  It is conceivable.   At worst what 

I propose will be about the same strength as now.  Let's face it, PxQ recapture often is not terribly deep

and does not require the full brainpower of leela.  A lot of moves inside the search tree are like that.

 

So no problem here I suppose.
Endgame on the other hand is a true weakness.

--I like the whole idea of the "ender" project and in fact suggested something like it before on this forum, 

albeit I was generally denounced when I did so... until eventually the leela crowd came round to my way of

thinking and incorporated endgame tablebases, and then DK also embarked on the ender project. 

Ender also might be enough to bring leela up to stockfish level, and at worst will just leave it about

the same strength as now.