Chessqueen wrote: ↑Sat Jan 21, 2023 2:00 am
In most ending position specially 2 Knights + Bishop versus Rook, no matter if The Rook is playing the White side or the Black side, in most games the three minors pieces which is equal to 9.0 is superior to the 2 Rooks
Well People who have used Chessmaster GUI have probably been messing around with Piece values for around 30 years or so now because 'Chessmaster' had settings where you could adjust pawns queens..rooks...Bishops...etc for the various "personalities".....Most of the "settings" are variable depending on the game type you are playing....(Open or closed...Early or middlegame or Late) plus positional considerations...A pawn could be anwhere between 0 and 10 depending on where it is (and when)...A knight 0 and 8... two connected knights 6-10 two terrible knights could be worthless! A bad bishop could be a negative!...two good bishops worth almost a queen....Two Rooks are probably not normally 10.0 but sometimes could be more....A queen is always the most powerful piece on the board but only in concert with other pieces...(knight...pawns Rook Bishop king etc) A queen alone is worth only a rook & a Bishop...(6-7 points) Actual piece values are probably a "trade secret"... all the programmers use some number and probably they run the gamut based on what I've seen...Morphy (& Komodo + others) in their "Odds Games" gives a great hint on these issues....
ps...Thats why you have to be real...real careful playing "point count" chess....Two connected rooks alone would be unable to push one pawn past any squares controlled by the 3 minor pieces in concert...in close quarters the 3 minor pieces reign...in a long distance struggle perhaps the rooks might be enough (depending on tactical considerations)....3 minor pieces will prevail against a queen in my book (tactics permitting)....
Chessqueen wrote: ↑Sat Jan 21, 2023 2:00 am
In most ending position specially 2 Knights + Bishop versus Rook, no matter if The Rook is playing the White side or the Black side, in most games the three minors pieces which is equal to 9.0 is superior to the 2 Rooks
ps...Thats why you have to be real...real careful playing "point count" chess....in a long distance struggle perhaps the 2rooks might be enough against 3 minor pieces(depending on tactical considerations)....3 minor pieces will prevail against a queen in my book (tactics permitting)....
Can you provide 2 Samples of those positions that you have in mind, and let top 4 engine solve it
Funny, because that is something I find too from material imbalance games from opening positions with Fairy-Max. Rooks consistently test about 25 cP lower than yyou would expect. The same holds for other pieces with only orthogonal moves, such as the Wazir. If you start that before the Pawns it tests as about 140cP (slightly below Ferz), but starting behind the Pawn wall it is only 120 cP.
I think this has to do with open-file bonus; an orthogonal piece is quite difficult to develop if it starts behind a closed rank of FIDE Pawns. But of course in the end-game most files will be open, and the classical value of the Rook might thus incorporate an open-file bonus.
These posts are mostly nonsense and I doubt anyone pays much mind or interest. I recollect posting the definitive analysis of relative piece values several years ago and nobody took any notice of that either.
Simple method:
Take your chess engine of choice, autoplay several tens of thousand of games at fast time controls (to ensure a reasonable number of non-draws).
Generate the resulting EPDs for all game positions.
Split the EPDs into about eight or so buckets, depending on piece count remaining.
For each bucket (should have a few hundred thousand positions) perform linear regression analysis, treating the data as a linear polynomial of weight x count of piece type against game result.
This will give you a graph of relative value (for each piece) against game phase (the buckets). Obviously normalising the relative values is useful.
Hey presto, if you used SF NNUE, you’ll probably end up with some good figures and then these threads can get pointed at the definitive results.
It’s also v v useful as an engine comparator. Similar engines produce similar shaped graphs (value against phase) which is very hard to disguise.
Edit: possibly the comparator and the relative results would be even better if EPD best line value was used instead of game result. Actually I forget which I used at the time.
chrisw wrote: ↑Wed Nov 26, 2025 3:00 pm
Simple method:
Take your chess engine of choice, autoplay several tens of thousand of games at fast time controls (to ensure a reasonable number of non-draws).
Generate the resulting EPDs for all game positions.
Split the EPDs into about eight or so buckets, depending on piece count remaining.
For each bucket (should have a few hundred thousand positions) perform linear regression analysis, treating the data as a linear polynomial of weight x count of piece type against game result.
This will give you a graph of relative value (for each piece) against game phase (the buckets). Obviously normalising the relative values is useful.
Agreed, that is a valid approach which should be quite informative. Evolving piece values with game phase is an enduring topic. Have you presented your results anywhere? I found a Sun Apr 16, 2023 mention here without follow-up.
https://talkchess.com/viewtopic.php?p= ... n#p946426
But there, it was another member, ferdy, who shared his numbers and plots. After venturing off original topic, as discussion board threads tend to do, but beneficially so.
chrisw wrote: ↑Wed Nov 26, 2025 3:00 pm
Simple method:
Take your chess engine of choice, autoplay several tens of thousand of games at fast time controls (to ensure a reasonable number of non-draws).
Generate the resulting EPDs for all game positions.
Split the EPDs into about eight or so buckets, depending on piece count remaining.
For each bucket (should have a few hundred thousand positions) perform linear regression analysis, treating the data as a linear polynomial of weight x count of piece type against game result.
This will give you a graph of relative value (for each piece) against game phase (the buckets). Obviously normalising the relative values is useful.
Agreed, that is a valid approach which should be quite informative. Evolving piece values with game phase is an enduring topic. Have you presented your results anywhere? I found a Sun Apr 16, 2023 mention here without follow-up.
https://talkchess.com/viewtopic.php?p= ... n#p946426
But there, it was another member, ferdy, who shared his numbers and plots. After venturing off original topic, as discussion board threads tend to do, but beneficially so.
Yes, somewhere, but nothing durable. Might have another crack at it at some point, the graph shapes are a signature so useful for comparing allegedly different engines/nets/training
