Why does the computer promote to rook?

when two moves are totally equal some program pick randomly from equal moves and most them pick one that on top list of moves to investigate.

@petri999 said in #11:

when two moves are totally equal some program pick randomly from equal moves and most them pick one that on top list of moves to investigate.

This brings about the notion of separation power of the engine microscope. I would rephrase, "when 2 moves are scored different only below separation ability of the engine given its assumptions/valuation on non-terminal outcome positions information", then what you said.

They are equal within engine programmed way of measuring position information as a score toward chess objectives which are in the legal set of rules, or any held belief as to what part of the position information is relevant to predict favored outcome (winning). So obviously, the well known most recently used 1,3,3,5,9 (and ? for King) material count system summarizing the power of each class of mobility rules (one per chess unit material type) (point: it is a thumb rule reflecting the legal differences in mobility for each piece.

Then one has to deal with the logical consequences of making the bet that this will be enough. This selective view of what part of the full position information will be scored, and which positions in tree explored will receive the right to even be score that way.

One choice in what is information in a position, has consequences in which positions get to be measured, and then the whole development might become prisoner of those initial assumptions. The material count system not enough when tree sizes grow big with faster machines (more terminal nodes mixed in), and even putting some positional labels on extra evaluation terms, might not fix the neglect of part of position information because the consequences on tree search design would anyway make that apparently more complete information measure only be used on certain positions selected per characteristics consequent to initial bias or incomplete position information.

A reason for that is that parameters of old for the material thumb rules are fixed, and the new terms would be optimized with such original constraints with the result that NNue had to come rescue for such blind spot (to a certain extent "of moderate depth" before conversion (of "small" imbalance at root) into material count visible position information could be seen).

The new positional parameters meant to compensate for the only material count partial use of position information, are still being optimized with the same tree search organization that came as tandem consequence to not having to need the extra information. meaning, that in the tree search design of which nodes get evaluation, there might be limited support for the true parameters that would be optimal if every nodes were to get a static evaluation (as in a0 and lc0 do, actually).

But interesting observation from the op. We should accumulate analytic "anecdotes" about engine analysis, for eventually making the case for engine being better analytical tools, which might have some ELO cost (or not, but likely) by finding patterns of current engine scoring that seem odd from human chess point of view.

@GoneSoul said in #1:
There is no reason, it doesn't really matter as the piece will be captured in the next move

@Elite_Soul said in #4:

If we really are using this logic,then why isnt it promoting to a bishop or a knight instead?
Because it wants to check as part of the puzzle

@GoneSoul said in #1:

I've noticed that in a lot of puzzles where the opponent promotes but then loses the promoted piece right away, the piece promotes to a rook? Why not promote to a queen? I know it doesn't matter since the piece is lost anyway, it just seems strange to me.

Here is an example of what I'm talking about:
lichess.org/training/hangingPiece/xqnXf

From a practical stand point. Does it matter? The piece of choice is irrelevant as it is going to get captured regardless which one is it. However, the puzzles are taken from human games, so the guy playing that move chose the rook. Dont ask me why, but my guess is that its faster if you dont have autoqueen. The rook is exactly where the cursor is, the queen is above and bishop and knight are below. So double click promotes to a rook.

A computer might as well choose any piece from a technical stand point if the piece had no impact on the position, it would probably pick the first one coded.

Prolly a rookie!?

It is random

@GoneSoul said in #1:

I've noticed that in a lot of puzzles where the opponent promotes but then loses the promoted piece right away, the piece promotes to a rook? Why not promote to a queen? I know it doesn't matter since the piece is lost anyway, it just seems strange to me.

Here is an example of what I'm talking about:
lichess.org/training/hangingPiece/xqnXf

I think, it was answered

The computer looks to lose less material (bishop/knight, rook, queen), but still with the possibility of giving mate (rook, queen), and if there is a possible check as well. The common factor in this case: rook. So the computer promotes to a rook.