- Blind mode tutorial
lichess.org
Donate

Science of Chess: What makes a move seem "Brilliant?"

@NDpatzer said in #3:

Ack - good catch, fixing it now! Thanks!

@NDpatzer said in #3: > Ack - good catch, fixing it now! Thanks!

In my mind, an automated brilliant/great move should be some linear combination of the following:

  1. The "stability" of the position before the brilliant move. This is a measure of how easily the evaluation changes depending on the move played.
  2. The analysis depth required before the move played appears good.

Under this definition, immediate sacrifices are the most common type of great move, but not only kind. However, all great moves by this definition are mathematically a type of "sacrifice".
Aesthetics are lacking in this definition, but trying to pin down what's "not obvious to a human" requires extra model to tack on beyond what Stockfish comes with. I'd rather let the computer show me the mathematically "great" moves and then I can decide which ones are truly great from that larger subset. Adding rules of what humans think is a great move may actually be like looking backwards at our own intelligence about chess rather than opening our minds and becoming more capable and appreciative players.

In my mind, an automated brilliant/great move should be some linear combination of the following: 1. The "stability" of the position before the brilliant move. This is a measure of how easily the evaluation changes depending on the move played. 2. The analysis depth required before the move played appears good. Under this definition, immediate sacrifices are the most common type of great move, but not only kind. However, all great moves by this definition are mathematically a type of "sacrifice". Aesthetics are lacking in this definition, but trying to pin down what's "not obvious to a human" requires extra model to tack on beyond what Stockfish comes with. I'd rather let the computer show me the mathematically "great" moves and then I can decide which ones are truly great from that larger subset. Adding rules of what humans think is a great move may actually be like looking backwards at our own intelligence about chess rather than opening our minds and becoming more capable and appreciative players.

If chess com says it’s brilliant it must be brilliant

If chess com says it’s brilliant it must be brilliant

I got interested in this subject when I tried to fix an old program. This old program already makes automated annotations based on a min-ply and max-ply analysis. So its based on science, AND its not made for any marketing use or purpose. Its simply an attempt to capture the science of "what makes a move interesting from a human commenter point of view".

The implementation is based on an old science post in Cambridge. It has formulas based on a "shallow" search and a "deep" search done by a chess engine of choice. It checks the gradient from an obvious low-ply analysis against a deep max-ply one.
So yes, there are formulas here for finding non-obvious "not so good" moves that turn out to be good in a deeper max-ply analysis.

The implementation I am referring to is the github Picochess program that in addition to running a chess engine you play against it also runs an analysis engine that performs this min-ply-max-ply analyse and automatically annotates the user moves.
There are formulas for !!, !, !? and ?!

This would need to be updated as the science is 10 years (?) old and the program itself is more than 5 years old. I have documented the current formula and old Cambridge texts in a PDF in this discussion. And we would need to come up with possibly new or improved formulas,
https://github.com/JohanSjoblom/picochess/discussions/39
If someone has research or formulas on this topic I would be happy to take part in that.
BR, Johan Sjöblom, Vaasa Chess Club, Finland

I got interested in this subject when I tried to fix an old program. This old program already makes automated annotations based on a min-ply and max-ply analysis. So its based on science, AND its not made for any marketing use or purpose. Its simply an attempt to capture the science of "what makes a move interesting from a human commenter point of view". The implementation is based on an old science post in Cambridge. It has formulas based on a "shallow" search and a "deep" search done by a chess engine of choice. It checks the gradient from an obvious low-ply analysis against a deep max-ply one. So yes, there are formulas here for finding non-obvious "not so good" moves that turn out to be good in a deeper max-ply analysis. The implementation I am referring to is the github Picochess program that in addition to running a chess engine you play against it also runs an analysis engine that performs this min-ply-max-ply analyse and automatically annotates the user moves. There are formulas for !!, !, !? and ?! This would need to be updated as the science is 10 years (?) old and the program itself is more than 5 years old. I have documented the current formula and old Cambridge texts in a PDF in this discussion. And we would need to come up with possibly new or improved formulas, https://github.com/JohanSjoblom/picochess/discussions/39 If someone has research or formulas on this topic I would be happy to take part in that. BR, Johan Sjöblom, Vaasa Chess Club, Finland

@messier109 said in #14:

I got interested in this subject when I tried to fix an old program. This old program already makes automated annotations based on a min-ply and max-ply analysis. So its based on science, AND its not made for any marketing use or purpose. Its simply an attempt to capture the science of "what makes a move interesting from a human commenter point of view".

The implementation is based on an old science post in Cambridge. It has formulas based on a "shallow" search and a "deep" search done by a chess engine of choice. It checks the gradient from an obvious low-ply analysis against a deep max-ply one.
So yes, there are formulas here for finding non-obvious "not so good" moves that turn out to be good in a deeper max-ply analysis.

The implementation I am referring to is the github Picochess program that in addition to running a chess engine you play against it also runs an analysis engine that performs this min-ply-max-ply analyse and automatically annotates the user moves.
There are formulas for !!, !, !? and ?!

This would need to be updated as the science is 10 years (?) old and the program itself is more than 5 years old. I have documented the current formula and old Cambridge texts in a PDF in this discussion. And we would need to come up with possibly new or improved formulas,
github.com/JohanSjoblom/picochess/discussions/39
If someone has research or formulas on this topic I would be happy to take part in that.
BR, Johan Sjöblom, Vaasa Chess Club, Finland

Thanks for sharing this - this is really excellent! I'm hoping to get some students excited about chess research and this might be a nice start to talking about a project. I'll be in touch if anyone expresses interest, and hopefully some other folks here might be up for working on this too.

@messier109 said in #14: > I got interested in this subject when I tried to fix an old program. This old program already makes automated annotations based on a min-ply and max-ply analysis. So its based on science, AND its not made for any marketing use or purpose. Its simply an attempt to capture the science of "what makes a move interesting from a human commenter point of view". > > The implementation is based on an old science post in Cambridge. It has formulas based on a "shallow" search and a "deep" search done by a chess engine of choice. It checks the gradient from an obvious low-ply analysis against a deep max-ply one. > So yes, there are formulas here for finding non-obvious "not so good" moves that turn out to be good in a deeper max-ply analysis. > > The implementation I am referring to is the github Picochess program that in addition to running a chess engine you play against it also runs an analysis engine that performs this min-ply-max-ply analyse and automatically annotates the user moves. > There are formulas for !!, !, !? and ?! > > This would need to be updated as the science is 10 years (?) old and the program itself is more than 5 years old. I have documented the current formula and old Cambridge texts in a PDF in this discussion. And we would need to come up with possibly new or improved formulas, > github.com/JohanSjoblom/picochess/discussions/39 > If someone has research or formulas on this topic I would be happy to take part in that. > BR, Johan Sjöblom, Vaasa Chess Club, Finland Thanks for sharing this - this is really excellent! I'm hoping to get some students excited about chess research and this might be a nice start to talking about a project. I'll be in touch if anyone expresses interest, and hopefully some other folks here might be up for working on this too.

@NDpatzer said in #15:
I loaded the Fischer game into Picochess with the position just before Fischers move Nxf2.
The automated analysis result came out as follows below. Note: I have collected all formulas and Cambridge texts into a PDF that can be found in above github discussion: https://github.com/JohanSjoblom/picochess/discussions/39

In short the result for Nxf2 in current picochess is:
evaluated score for the move (depth 23) = 340
D2 / Delta 2 = (lost centipawns for the move) = 0
DS / Delta S = (gradient from low-ply 5 depth to max-ply 23 depth) = 340 - 88 = 252

Picochess at the moment only uses Delta2 and DeltaS. And for Delta2 the requirement for !! is zero and that is True in this case. The requirement for the gradient delta S currently has a parameter threshold of 350 for !! and 250 for ! so current Picochess only gave one ! as in good move for Fischers Nxf2.
One reason is that the low-ply of 5 depth already gives this move the highest low-ply score of 88, and so the difference to 340 is not large enough for the current parameters in Picochess. The move comes out as best move already in the low-ply 5 depth analysis.

In general, from a science and research point of view my 10cent opinion is this:

  • It is fairly easy to establish the formulas to use for evaluation
  • BUT: It is quite challenging to find a good parameter value set, like for instance the thresholds above for !! and ! are 350 and 250 respectively. A training set based on famous plays like Fischer could be established and based on those it could propose a good parameter setting to use.
@NDpatzer said in #15: I loaded the Fischer game into Picochess with the position just before Fischers move Nxf2. The automated analysis result came out as follows below. Note: I have collected all formulas and Cambridge texts into a PDF that can be found in above github discussion: https://github.com/JohanSjoblom/picochess/discussions/39 In short the result for Nxf2 in current picochess is: evaluated score for the move (depth 23) = 340 D2 / Delta 2 = (lost centipawns for the move) = 0 DS / Delta S = (gradient from low-ply 5 depth to max-ply 23 depth) = 340 - 88 = 252 Picochess at the moment only uses Delta2 and DeltaS. And for Delta2 the requirement for !! is zero and that is True in this case. The requirement for the gradient delta S currently has a parameter threshold of 350 for !! and 250 for ! so current Picochess only gave one ! as in good move for Fischers Nxf2. One reason is that the low-ply of 5 depth already gives this move the highest low-ply score of 88, and so the difference to 340 is not large enough for the current parameters in Picochess. The move comes out as best move already in the low-ply 5 depth analysis. In general, from a science and research point of view my 10cent opinion is this: - It is fairly easy to establish the formulas to use for evaluation - BUT: It is quite challenging to find a good parameter value set, like for instance the thresholds above for !! and ! are 350 and 250 respectively. A training set based on famous plays like Fischer could be established and based on those it could propose a good parameter setting to use.