underpromoting is worth how many elo?

[Warren D Smith]

If stockfish no longer thinks about nor performs underpromotions, then how many elo does that cost it versus

full stockfish?

Doesn't matter, of course, just a pure amusement/curiosity question.

I suppose we could ask fancier questions of this kind ("how many elo is it worth to know the rules?"), 

such as not knowing about specific kinds of underpromotions (knight, bishop, etc);

or not knowing about en passant.  (I actually played chess for several years before realizing you

were not allowed to castle when in check...)

[Warren D Smith]

On 3/8/20, Warren D Smith <warre...@gmail.com> wrote:
> If stockfish no longer thinks about nor performs underpromotions, then how
> many elo does that cost it versus
> full stockfish?
>
> Doesn't matter, of course, just a pure amusement/curiosity question.
>
> I suppose we could ask fancier questions of this kind ("how many elo is it
> worth to know the rules?"),
> such as not knowing about specific kinds of underpromotions (knight,
> bishop, etc);

--despite the humor, this actually is a serious question.

I claim it is entirely conceivable that stockfish and/or leela zero
would be (a) simpler code, and/or (b) measurably stronger elo, if they
never thought about underpromoting to a bishop or rook.

Which if so would be rather paradoxical. Mind you I am not recommending SF
go this route, but if that happens then SF (according to their own
stated policies about elo measurement and simplicity!) would do it
despite my feelings...

--
Warren D. Smith
http://RangeVoting.org <-- add your endorsement (by clicking
"endorse" as 1st step)

[marku.b...@gmail.com]

Think promotion should only have 2 options. Either Knight or Queen. Because Queen does everything a Bishop and Rook does, but in case of a rare type of mate where a knight promotion would secure it, if you promote a queen it might go into stalemate or delay of mate

[quant...@gmail.com]

On Sunday, March 8, 2020 at 6:25:04 PM UTC-5, marku....@gmail.com wrote:

Think promotion should only have 2 options. Either Knight or Queen. Because Queen does everything a Bishop and Rook does, but in case of a rare type of mate where a knight promotion would secure it, if you promote a queen it might go into stalemate or delay of mate

If you're seeking *correctness* you do need all types of promotions. There exist situations where promoting to a queen leads to stalemate while promoting to a bishop leads to a checkmate. Contrived? Sure, but they do exist.

[Warren D Smith]

> If you're seeking *correctness* you do need all types of promotions. There
> exist situations where promoting to a queen leads to stalemate while
> promoting to a bishop leads to a checkmate. Contrived? Sure, but they do
> exist.

--yes, and I feel the "world's best chess program" (WBCP) ought to be
good at solving
problems, and constructed problems often include underpromotions in
the solution.
Most famously the "Babson task" problem, which is an amazing feat of
chess construction.
Similarly I feel WBCP ought to be able to mate with bishop & knight.

However, probably if SF were unable to mate with B&N, or unable to
promote to B&R, it would still have won & drawn every tournament game
it ever won/drew. The impact of these things on its strength is
small, I'd guess below 1 elo. And in the case of underpromoting to
B&R, that actually complicates and slows move representation, move
generation and search, which perhaps actually hurts SF's strength by
an amount greater than the benefit from knowing the correct rule!

Which, if so, would at least be amusing to know. Also, it has been
suggested that underpromoting to knight usually happens only when that
causes a check. If we ignored
N-underpromos that do not check, the impact on strength again
conceivably might be small enough that it again would be smaller than
the speed benefit from reducing the search...

[Warren D Smith]

Actually, these things might be too small for FishTest to measure in a
reasonable amount of
time.

Which leads to another interesting and perhaps valuable idea:
"indirect measurement"!
I claim it is possible to measure some things much more accurately
than is feasible with FishTest.

1. we can measure speed very precisely, certainly 1 part per million,
probably billion. We know a relationship between elo and speed:
specifically, about 70 elo points per speed-doubling, i.e. about 1 elo
per 1% speed increase.

2. We can measure frequency of occurrence of various phenomena (such
as underpromotions, KBNk endings) pretty precisely. KBNk ending
allegedly occurs
1/5000 fraction of games. Underpromotions allegedly occur in about
1/1200 of games.

These numbers all are not due to me and I do not vouch for their accuracy.

So if SF could not mate with bishop & knight, that'd hurt it by about 1/10000 of
a full point per game, i.e. about 0.1 elo. If SF could not
underpromote (all types)
that would hurt it by somewhere around 0.5 elo. The speed benefit
from SF ignoring
all underpromotions would outweigh that if it were >0.5% speed gain.
We could measure the speed gain to much better accuracy than that.
I suspect about 0.1 to 1% of all searched moves are pawn-to-8th, and
hence the slowdown likely is of the correct order of magnitude to
outweigh underpromotion benefits, especially
if we only are talking about R&B underpromos.

So, this rough analysis has suggested, paradoxically, that knowing the
rules actually hurts stockfish. But the hurt is at most a few elo.








If a modification to SF impacts speed

[protonspring]

I think I tested this a few years ago. . . when I removed everything but queen, it would not pass (showed some regression).

[Mindbreaker]

I would suggest the rule: If promoting to a Queen or Knight results in a stalemate or us getting checkmated in 6 plys or less...then look at Bishop or Rook promotion.