Your network blocks the Lichess assets!

lichess.org
Donate

Identifying Brilliant Moves With Leela Chess Zero

@jk_182

Yes I understand, using policy of one node (pairs) is about the bulk odds, while the more tactical turn by turn would be uncovered with more forward scrutiny on policy breadth, so on the other hand relying on SF for its exhaustive contrasting turn by turn view (hoping it is not too blinded, but nobody is really asking, about things not deeper deeper).

Which makes me wonder about what is human brilliant call. Is it always about a short term upheaval, or could it be a long term too (in the human set of calling outs in dataset, are they having some clustering of their own, again with the other angle, but I think your dimensions of surprise and strength as long as keeping their goggle angle explicit can serve to charaterize such subets of human called brilliant position pairs (with possible variance in depth of visible reward that would make the call by experts in hindsight possible). It may need other parameters of analysis from LC0 and SF as well.. (such as SF discovery depth from brilliant moves, might need looking inside the subtree searched entrails suring iterative deepening)... other things you might want to explore.

but I get the policy at one node which is basically about global odds evaluation more than tactical percision. and so might represent the more intuitive or expected moves. Lc0 has had its share of blindspot tactical pockets..

Also, maybe the phase of brilliant moves (human or 2-type machine candidate definitions), as LC0 global odds are very good at early phase, but as it exploits more than explores in endgame territory, it weak tactical pockets there might be more present (or prevalent, but in what sense, we would need some sense of distance, yes, to invoke pockets and what would prevalence be, if even human players don't wander much).

@jk_182 Yes I understand, using policy of one node (pairs) is about the bulk odds, while the more tactical turn by turn would be uncovered with more forward scrutiny on policy breadth, so on the other hand relying on SF for its exhaustive contrasting turn by turn view (hoping it is not too blinded, but nobody is really asking, about things not deeper deeper). Which makes me wonder about what is human brilliant call. Is it always about a short term upheaval, or could it be a long term too (in the human set of calling outs in dataset, are they having some clustering of their own, again with the other angle, but I think your dimensions of surprise and strength as long as keeping their goggle angle explicit can serve to charaterize such subets of human called brilliant position pairs (with possible variance in depth of visible reward that would make the call by experts in hindsight possible). It may need other parameters of analysis from LC0 and SF as well.. (such as SF discovery depth from brilliant moves, might need looking inside the subtree searched entrails suring iterative deepening)... other things you might want to explore. but I get the policy at one node which is basically about global odds evaluation more than tactical percision. and so might represent the more intuitive or expected moves. Lc0 has had its share of blindspot tactical pockets.. Also, maybe the phase of brilliant moves (human or 2-type machine candidate definitions), as LC0 global odds are very good at early phase, but as it exploits more than explores in endgame territory, it weak tactical pockets there might be more present (or prevalent, but in what sense, we would need some sense of distance, yes, to invoke pockets and what would prevalence be, if even human players don't wander much).

It's also brilliant when player that is playing a losing game finds a move that creates a fortress or finds a way to draw the game. Many player's want respect when they are winning in the last phase of the game, but some want it immediately after the opening. The opponent might be playing a losing game, but the game is not over until it's obvious that there is no hope to draw. Completing a match is not poor sportsmanship, and it's a chance to see the power a GM has and how it would humanly finish the game. Resigning without reaching the middle game seems hasty and the database most likely has many games that end hastily. I believe those games should be filtered out before using the rest of the games for NN or WDL % ratios.

The brilliancy of moves should be visible in a graph, like a change in directional flow of the graph. The brilliant move does not need to shift the graph from negative to positive values.

Devise a simple solution to a complex problem:
The inverse of a brilliant move, is a blunder.
The inverse of a good move, is an error.
The inverse of an accurate move, is an inaccuracy.

A brilliant choice just simplified a situation that was complex. A creative choice is brilliant when you need to simplify or make the game complex to overcome a losing game. We maintain the pressure, when we do not have the advantage needed to win. We simplify when the timing is ripe. If an engine innovates compared to the other engines, than that's brilliant too.

A "Brilliant" move is a "Smart" move, because the timing for the move is dead on. The move could not fall at a better time.
A brilliant move is not only smart, but also well-timed !

Thinking out of the box for inspiration, to make a formula for chess brilliance.
SMART formula for setting goals stands for Specific, Measurable, Achievable, Relevant, and Time-Bound.
https://us.humankinetics.com/blogs/excerpt/5-rules-for-setting-smart-goals

It's also brilliant when player that is playing a losing game finds a move that creates a fortress or finds a way to draw the game. Many player's want respect when they are winning in the last phase of the game, but some want it immediately after the opening. The opponent might be playing a losing game, but the game is not over until it's obvious that there is no hope to draw. Completing a match is not poor sportsmanship, and it's a chance to see the power a GM has and how it would humanly finish the game. Resigning without reaching the middle game seems hasty and the database most likely has many games that end hastily. I believe those games should be filtered out before using the rest of the games for NN or WDL % ratios. The brilliancy of moves should be visible in a graph, like a change in directional flow of the graph. The brilliant move does not need to shift the graph from negative to positive values. Devise a simple solution to a complex problem: The inverse of a brilliant move, is a blunder. The inverse of a good move, is an error. The inverse of an accurate move, is an inaccuracy. A brilliant choice just simplified a situation that was complex. A creative choice is brilliant when you need to simplify or make the game complex to overcome a losing game. We maintain the pressure, when we do not have the advantage needed to win. We simplify when the timing is ripe. If an engine innovates compared to the other engines, than that's brilliant too. A "Brilliant" move is a "Smart" move, because the timing for the move is dead on. The move could not fall at a better time. A brilliant move is not only smart, but also well-timed ! Thinking out of the box for inspiration, to make a formula for chess brilliance. SMART formula for setting goals stands for Specific, Measurable, Achievable, Relevant, and Time-Bound. https://us.humankinetics.com/blogs/excerpt/5-rules-for-setting-smart-goals

@jk_182 said in #13:

It's not quite finding moves where SF and LC0 don't agree, since only the policy of LC0 is used. The idea is that LC0 views the position without any calculation to determine which moves are surprising. What I didn't want to happen is that a basic queen sacrifice leading to a back rank mate will be called brilliant (which in reality it isn't since every relatively strong player sees the idea instantly) and LC0 on one node is strong enough find common combinations so they have a higher policy. SF is only used to determine the objective quality of the move since many surprising moves are simply bad (for example, 1.e4 e5 2.f4 has a policy of less than 1 percent).

I didn't test it too deeply since it took a lot of time, but I started with known brilliant moves to see if the idea works for these examples. Then I looked at some more games to see if there aren't too many brilliant moves. I hope this answers your question.

I didn't use the sharpness, since I was also interested about great positional moves (I think that Ra2 in the Petrosian-Olafsson game is a good example of that). But using sharpness as an additional indicator can give you more info about the type of position.

I haven't considered using the whole probability distribution, but I'm not quite sure whether a move is more brilliant if there is one "obvious" alternative which is worse or if there are many "obvious" moves but none as strong as the strongest move with a low policy. I think the best way to overcome the issue with many good options is to say that a brilliant move has to be clearly stronger than any other move. I haven't added this since I this would have doubled the computing time and I could only run it on my laptop.

I took time to read more.. And you actually did explain a lot more your process and reasoning. thanks.
I am glad you are actually keeping a question period. That is where things are best explained finally.

I did not pay attention, is there some place we can see the details of the work. Is that from the links in the external journal site? (going back to blog.. to check is I skipped something).

@jk_182 said in #13: > It's not quite finding moves where SF and LC0 don't agree, since only the policy of LC0 is used. The idea is that LC0 views the position without any calculation to determine which moves are surprising. What I didn't want to happen is that a basic queen sacrifice leading to a back rank mate will be called brilliant (which in reality it isn't since every relatively strong player sees the idea instantly) and LC0 on one node is strong enough find common combinations so they have a higher policy. SF is only used to determine the objective quality of the move since many surprising moves are simply bad (for example, 1.e4 e5 2.f4 has a policy of less than 1 percent). > > I didn't test it too deeply since it took a lot of time, but I started with known brilliant moves to see if the idea works for these examples. Then I looked at some more games to see if there aren't too many brilliant moves. I hope this answers your question. > > I didn't use the sharpness, since I was also interested about great positional moves (I think that Ra2 in the Petrosian-Olafsson game is a good example of that). But using sharpness as an additional indicator can give you more info about the type of position. > > I haven't considered using the whole probability distribution, but I'm not quite sure whether a move is more brilliant if there is one "obvious" alternative which is worse or if there are many "obvious" moves but none as strong as the strongest move with a low policy. I think the best way to overcome the issue with many good options is to say that a brilliant move has to be clearly stronger than any other move. I haven't added this since I this would have doubled the computing time and I could only run it on my laptop. I took time to read more.. And you actually did explain a lot more your process and reasoning. thanks. I am glad you are actually keeping a question period. That is where things are best explained finally. I did not pay attention, is there some place we can see the details of the work. Is that from the links in the external journal site? (going back to blog.. to check is I skipped something).

@Toscani said in #17:

Found some stuff about brilliancies and narrowness.

thanks for those links. Narrowness as natural language, does seem similar to sharpness. Maybe a way to avoid stepping on existing concepts while developing more automatic or covering definitions.

I also like it, but sharpness too, for its spatial implied internal model. Narrow where or how.. Is there is a distance, that could measure that?

Are we talking strictly about laying some SF PV profile forced to be all the legal moves from a position. putting the highest score at center of some graph, and then ordering down the others per some fixed arbitrary abscissa tick marks? (for more visual impart spread those one each side, and then looking at slope at maximal score.

which can also be done for LC0 in one node policy, I would think. That is what I meant by distribution. The shape of it. But if the position had only one good move, and all others bad, which I associate to sharp, it might not be very brilliancy prone position. (it depends on whether that move spike is difficult to find? I get confused).

I wonder if it did it not use to be called zero policy, or 0 node policy. I guess depends where the node index starts.
My understanding here, is one node from root, node being the position after move applied to root, just to be annoying with blabbering precision, never who might be reading with what definition of node in some decision tree, or state space graph from which to make decisions). So basically the policy probability is the successor node evaluation relative to all other evaluations. Unlike SF, my understanding is that LC0 has an evaluation function for all the possibly nodes. depth or width are potentially all equivalent improving ways to gain some breadth based increased "accuracy" on move potential reward. But it is not about 2 engines disagreeing on best, but I think, exploiting the differences (well it seems only LC0 knobs being considered) in scoring search control parameters.

The fact that LC0 can attribute to each immediate successors without any concern on its quiescence status some evaluation of its zero node odds (WDL or some reward normal form version of that 2 dimensional probability, right, either draw or W or lose making 100%, so only need 2 odds, maybe I should read the sigmoid formulas....:) ). Is what is being leveraged here..

And I understand that @jk_182 from last reply to mine, was already looking for long term, or positional distinction among human brilliancy labelled moves at their root positions (a pair, either root and move, or root and successor).

So, what I said about SF entrails also being possible knobs to play with, to depth of discovery, might be something that could help with the term of the leaf evaluation discovery within the tree. I tend to associate positional good moves with restrictions on the range of moves in descendant position futures, so long term effects. But we are also in some drifting ground here. Positional is just not immediate material changing move. It can apply to tactical term or longer term. (here again, made an association between tactical and short term, debatable? the patterns of position feature signals and action patterns are of short term natures, why puzzles seem sufficient to explore elemental isolated such patterns).

But my initial understanding of LC0 weaknesses about best move, was wrong, to some extent. It is not clear the trade off being made and the interpretation but there may be room as the author suggests with this work, while zero policy (i.e. one node evaluation makes the policy) which is known, at least very early phase to be already containing lots of information from all the training games (the shallower the position, the more game outcomes would have to be statistically contributing to that position evaluation, if I understood the basic LC0 and A0 scheme of RL batches training schedule, from wide exploration to narrower but more informed exploitation of what was learned in previous batch).

Why I suggested or asked about the if that was a variable in the study of human called brilliancy annotations.

I am actually not just babbling in vain, I am testing if my understanding still stands while I am discussing. So feel free to contradict or adjust that understand, as you have done so far.

@Toscani said in #17: > Found some stuff about brilliancies and narrowness. thanks for those links. Narrowness as natural language, does seem similar to sharpness. Maybe a way to avoid stepping on existing concepts while developing more automatic or covering definitions. I also like it, but sharpness too, for its spatial implied internal model. Narrow where or how.. Is there is a distance, that could measure that? Are we talking strictly about laying some SF PV profile forced to be all the legal moves from a position. putting the highest score at center of some graph, and then ordering down the others per some fixed arbitrary abscissa tick marks? (for more visual impart spread those one each side, and then looking at slope at maximal score. which can also be done for LC0 in one node policy, I would think. That is what I meant by distribution. The shape of it. But if the position had only one good move, and all others bad, which I associate to sharp, it might not be very brilliancy prone position. (it depends on whether that move spike is difficult to find? I get confused). I wonder if it did it not use to be called zero policy, or 0 node policy. I guess depends where the node index starts. My understanding here, is one node from root, node being the position after move applied to root, just to be annoying with blabbering precision, never who might be reading with what definition of node in some decision tree, or state space graph from which to make decisions). So basically the policy probability is the successor node evaluation relative to all other evaluations. Unlike SF, my understanding is that LC0 has an evaluation function for all the possibly nodes. depth or width are potentially all equivalent improving ways to gain some breadth based increased "accuracy" on move potential reward. But it is not about 2 engines disagreeing on best, but I think, exploiting the differences (well it seems only LC0 knobs being considered) in scoring search control parameters. The fact that LC0 can attribute to each immediate successors without any concern on its quiescence status some evaluation of its zero node odds (WDL or some reward normal form version of that 2 dimensional probability, right, either draw or W or lose making 100%, so only need 2 odds, maybe I should read the sigmoid formulas....:) ). Is what is being leveraged here.. And I understand that @jk_182 from last reply to mine, was already looking for long term, or positional distinction among human brilliancy labelled moves at their root positions (a pair, either root and move, or root and successor). So, what I said about SF entrails also being possible knobs to play with, to depth of discovery, might be something that could help with the term of the leaf evaluation discovery within the tree. I tend to associate positional good moves with restrictions on the range of moves in descendant position futures, so long term effects. But we are also in some drifting ground here. Positional is just not immediate material changing move. It can apply to tactical term or longer term. (here again, made an association between tactical and short term, debatable? the patterns of position feature signals and action patterns are of short term natures, why puzzles seem sufficient to explore elemental isolated such patterns). But my initial understanding of LC0 weaknesses about best move, was wrong, to some extent. It is not clear the trade off being made and the interpretation but there may be room as the author suggests with this work, while zero policy (i.e. one node evaluation makes the policy) which is known, at least very early phase to be already containing lots of information from all the training games (the shallower the position, the more game outcomes would have to be statistically contributing to that position evaluation, if I understood the basic LC0 and A0 scheme of RL batches training schedule, from wide exploration to narrower but more informed exploitation of what was learned in previous batch). Why I suggested or asked about the if that was a variable in the study of human called brilliancy annotations. I am actually not just babbling in vain, I am testing if my understanding still stands while I am discussing. So feel free to contradict or adjust that understand, as you have done so far.

other thought of possible knobs from LC0. When used with policy, typical the stopping criterion is about confidence interval to reach for distinguishing each candidate move from root as nodes set explored is increasing in breadth (how the nodes are being expanded is not something I fully understand**). I bet that does not really compute with existing UCI, so it have to be node count, and LC0 has some conversion curve to fit in. but the reasoning is that the node expansion is based on some confidence estimator being somehow available. Or what does the PUCT acronym really stand for? So, my suggestion, if I am not completely off base already, is that while using one node profiling of moves, maybe that current confidence interval among how moves successor evaluation differ is significant per such confidence estimator, would be an interesting variable.. (it might even have some human analogy, in that historically evolving knowledge, and perception or theory of the board, might also have been factors in evolution of what would have been called brilliancy.. I guess, given similar position, a previously known moves having been called brilliant once, won't be called brilliant anymore for all those sufficiently similar positions and that move.

** I did try, but those GO diagrams don't help me much somehow, or they are not at the mathematical level the A0 paper actually presented the whole framework, which I could understand there, in that state-action model formulation. it was so clear. It can also be my lack of satisfaction a game tree being able to completely describe all the information of the game world (all the games, that contribute to all the node evaluation, where being part of a pair of players solo game tree of decisions to consider, or all the possible pair of players decisions tree combined). my problem.. one day I will surrender to game trees being all there is... for now I would rather start with the state space and the action Siamese space (it can be described in the state space), where all pairs of players of the universe would have to navigate, even if each have their own game trees, which in their instances once cranked in depth, don't care about other branches nodes from other paths.. (yet the evaluation backprop. of outcome data, well it cares about all the paths taken from the position being evaluated). just spilling my current guts about my fog.

other thought of possible knobs from LC0. When used with policy, typical the stopping criterion is about confidence interval to reach for distinguishing each candidate move from root as nodes set explored is increasing in breadth (how the nodes are being expanded is not something I fully understand**). I bet that does not really compute with existing UCI, so it have to be node count, and LC0 has some conversion curve to fit in. but the reasoning is that the node expansion is based on some confidence estimator being somehow available. Or what does the PUCT acronym really stand for? So, my suggestion, if I am not completely off base already, is that while using one node profiling of moves, maybe that current confidence interval among how moves successor evaluation differ is significant per such confidence estimator, would be an interesting variable.. (it might even have some human analogy, in that historically evolving knowledge, and perception or theory of the board, might also have been factors in evolution of what would have been called brilliancy.. I guess, given similar position, a previously known moves having been called brilliant once, won't be called brilliant anymore for all those sufficiently similar positions and that move. ** I did try, but those GO diagrams don't help me much somehow, or they are not at the mathematical level the A0 paper actually presented the whole framework, which I could understand there, in that state-action model formulation. it was so clear. It can also be my lack of satisfaction a game tree being able to completely describe all the information of the game world (all the games, that contribute to all the node evaluation, where being part of a pair of players solo game tree of decisions to consider, or all the possible pair of players decisions tree combined). my problem.. one day I will surrender to game trees being all there is... for now I would rather start with the state space and the action Siamese space (it can be described in the state space), where all pairs of players of the universe would have to navigate, even if each have their own game trees, which in their instances once cranked in depth, don't care about other branches nodes from other paths.. (yet the evaluation backprop. of outcome data, well it cares about all the paths taken from the position being evaluated). just spilling my current guts about my fog.

The problem with using engines to determine when a move is brilliant has been shown by the hundreds of short videos that have flooded the internet since chess.com implemented it into their game analysis. It is simply impossible whether a move is actually brilliant by looking at its computer evaluation. There are two problems: Firstly the computer attaches "!!" too easily. Secondly, the computer does not know if the player playing the move actually understand why it is brilliant.

Of course, from a marketing perspective this makes sense. When your demographic is filled with players under a certain knowledge threshold it makes sense to lower the requirements for getting a "!" or "!!" move. That way, a largest part of your player base can get a sense of accomplishment that will make them want to come back for more.

In the end, whether a move is brilliant or not is not only based on the player looking at it, but also at their tactical, strategical and positional knowledge.

A good example is the mainlines of the King's Indian with 9. Ne1. When at some point ... Bxh3 comes whether you find it brilliant or not is based on your familiarity with the position. Player A that just started playing the KID might find it brilliant, Player B that has been playing mainline d4 theory finds it nice but typical and Player C who is experienced in the Mar de Plata structures might not even believe it is even worth discussing.

It is also a matter of when was the move played and when was it analyzed. Tal's Nd5, Ne6 and Nf5 sacrifices in the Sicilian could have been considered brilliant when he played them but know every slightly above average player knows about them. In 1950 Nd5 in the Sicilian was brilliant. In 2023 it is just typical.

Lastly, a big problem is that while many moves can be considered brilliant but the reason for them being so differs from move to move. Some may be tactically brilliant , meaning that the move is unexpected but it forcefully leads to some gain, and some may be strategically brilliant, meaning the move gives the player a good position even though nothing is forced.

For these reasons, I believe that a computer cannot be used to determine whether a move is brilliant or not.

The problem with using engines to determine when a move is brilliant has been shown by the hundreds of short videos that have flooded the internet since chess.com implemented it into their game analysis. It is simply impossible whether a move is actually brilliant by looking at its computer evaluation. There are two problems: Firstly the computer attaches "!!" too easily. Secondly, the computer does not know if the player playing the move actually understand why it is brilliant. Of course, from a marketing perspective this makes sense. When your demographic is filled with players under a certain knowledge threshold it makes sense to lower the requirements for getting a "!" or "!!" move. That way, a largest part of your player base can get a sense of accomplishment that will make them want to come back for more. In the end, whether a move is brilliant or not is not only based on the player looking at it, but also at their tactical, strategical and positional knowledge. A good example is the mainlines of the King's Indian with 9. Ne1. When at some point ... Bxh3 comes whether you find it brilliant or not is based on your familiarity with the position. Player A that just started playing the KID might find it brilliant, Player B that has been playing mainline d4 theory finds it nice but typical and Player C who is experienced in the Mar de Plata structures might not even believe it is even worth discussing. It is also a matter of when was the move played and when was it analyzed. Tal's Nd5, Ne6 and Nf5 sacrifices in the Sicilian could have been considered brilliant when he played them but know every slightly above average player knows about them. In 1950 Nd5 in the Sicilian was brilliant. In 2023 it is just typical. Lastly, a big problem is that while many moves can be considered brilliant but the reason for them being so differs from move to move. Some may be tactically brilliant , meaning that the move is unexpected but it forcefully leads to some gain, and some may be strategically brilliant, meaning the move gives the player a good position even though nothing is forced. For these reasons, I believe that a computer cannot be used to determine whether a move is brilliant or not.

@EphemeralAdvantage , doesn't that apply to ANY analysis, including human one? GothamChess can cry THE ROOOOK all he wants, there will be some GM somewhere scoffing because it was an obvious move. Or some noob with SF on, believing they understood the move.

I think the requirement to definitively determine if a move is brilliant or not is the chess equivalent of "perfect is the enemy of good".

There will be a number of attributes of a move, the purpose of these discussions is to eliminate the useless ones and focus on the more relevant. The impossibility of an objectively brilliant move doesn't matter If leela or maia can point us in the right direction.

@EphemeralAdvantage , doesn't that apply to ANY analysis, including human one? GothamChess can cry THE ROOOOK all he wants, there will be some GM somewhere scoffing because it was an obvious move. Or some noob with SF on, believing they understood the move. I think the requirement to definitively determine if a move is brilliant or not is the chess equivalent of "perfect is the enemy of good". There will be a number of attributes of a move, the purpose of these discussions is to eliminate the useless ones and focus on the more relevant. The impossibility of an objectively brilliant move doesn't matter If leela or maia can point us in the right direction.

@TotalNoob69 There is a difference between Gotham screaming about rook sacrifices and using a notation that has been part of chess since the first Informator. One is obviously a joke, they other tries to conceal itself as something serious. In the end, it is a marketing stunt to keep engagement, as I said above

"doesn't that apply to ANY analysis" Obviously it does. There are books for amateurs, experts, GMs and everything in between. Do you believe that a game analyzed both in a Chernev book and an Aagaard book will have the same notation? In one of his books Chernev put a "!" to 2. Nf3! after 1. e4 e5. It is all subjective.

"is to eliminate the useless ones" The fact that Leela thinks one move is hard to find while others are is of absolutely no importance. Marshall's ... Qg3 is spotted by modern engines in hundredths of a second, Botvinnik's Ba3 probably takes close to a tenth. Does this make these moves "useless" when searching for brilliant moves?

@TotalNoob69 There is a difference between Gotham screaming about rook sacrifices and using a notation that has been part of chess since the first Informator. One is obviously a joke, they other tries to conceal itself as something serious. In the end, it is a marketing stunt to keep engagement, as I said above "doesn't that apply to ANY analysis" Obviously it does. There are books for amateurs, experts, GMs and everything in between. Do you believe that a game analyzed both in a Chernev book and an Aagaard book will have the same notation? In one of his books Chernev put a "!" to 2. Nf3! after 1. e4 e5. It is all subjective. "is to eliminate the useless ones" The fact that Leela thinks one move is hard to find while others are is of absolutely no importance. Marshall's ... Qg3 is spotted by modern engines in hundredths of a second, Botvinnik's Ba3 probably takes close to a tenth. Does this make these moves "useless" when searching for brilliant moves?

I guess we're going back to "computers can't play chess", huh?

I guess we're going back to "computers can't play chess", huh?

If I understand it correctly, the tree in Leela Chess Zero (Lc0) expands with unknown (new FEN) positions and continues until a limit in the lc0-otions or the GUI is reached, such as Hash limits (limits the amount of data that can be processed). Increasing the Hash limit allows Lc0 to search more nodes (which I assume is better results). So the hardware ... like the clock cycle speed limits will impact the data getting processed (computing instructions per clock cycle) and so would the speed of the disk input/output (I/O). We buy what we can afford and more seems better than less. There is a balance to maintain in hardware parts to minimize bottlenecks. A smart move would be to buy the hardware that serves our own needs, depending on the chess engine we use. An old computer might not detect brilliant moves like a modern system. If I was to test bench for brilliant moves using the latest Lc0, than I would want a recent RTX30 or 40__ series. RAM is obviously needed for a chess engine to function adequately, and that's not counting how much the operating system is using or other programs running in the background. If my video card has 8G, than chances are I'd want 16G or more on my motherboard. Sorry got off topic.

Start with smaller datasets. They are quickly read from a drive and stored in ram. You don't want the dataset to end up in the swap file. If it does, performance goes down.

Depending on the GUI, a person can either set the engine at 1000 to 100,000 nodes per move or use centipawns to find brilliant moves. I guess it has to be a brilliant move when the move is rated higher than it is humanly possible.

https://en.wikipedia.org/wiki/Instructions_per_second
https://lczero.org/dev/wiki/lc0-options/
https://www.melonimarco.it/en/2021/03/08/stockfish-and-lc0-test-at-different-number-of-nodes/
https://www.gpucheck.com/gpu-benchmark-graphics-card-comparison-chart
https://www.cpuagent.com/
https://www.phoronix.com/news/LCZero-NVIDIA-Benchmarks
https://www.phoronix.com/review/nvidia-rtx3080-compute/6
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
https://www.cgdirector.com/best-computer-3d-modeling-rendering/
https://medium.com/the-mission/how-to-build-the-perfect-deep-learning-computer-and-save-thousands-of-dollars-9ec3b2eb4ce2

https://lucaschess.pythonanywhere.com/static/pdf/english/Lucas%20Chess%204-Tools.pdf

FEN (Forsyth–Edwards Notation)
GUI (Graphical User Interface)
PUCT (Predictor + Upper Confidence Bound tree search) algorithm.
SAN (Standard Algebraic Notation)
UCT (Upper Confidence Bound for Trees)

Has anyone else tried to find brilliant moves in popular opening games?

If I understand it correctly, the tree in Leela Chess Zero (Lc0) expands with unknown (new FEN) positions and continues until a limit in the lc0-otions or the GUI is reached, such as Hash limits (limits the amount of data that can be processed). Increasing the Hash limit allows Lc0 to search more nodes (which I assume is better results). So the hardware ... like the clock cycle speed limits will impact the data getting processed (computing instructions per clock cycle) and so would the speed of the disk input/output (I/O). We buy what we can afford and more seems better than less. There is a balance to maintain in hardware parts to minimize bottlenecks. A smart move would be to buy the hardware that serves our own needs, depending on the chess engine we use. An old computer might not detect brilliant moves like a modern system. If I was to test bench for brilliant moves using the latest Lc0, than I would want a recent RTX30 or 40__ series. RAM is obviously needed for a chess engine to function adequately, and that's not counting how much the operating system is using or other programs running in the background. If my video card has 8G, than chances are I'd want 16G or more on my motherboard. Sorry got off topic. Start with smaller datasets. They are quickly read from a drive and stored in ram. You don't want the dataset to end up in the swap file. If it does, performance goes down. Depending on the GUI, a person can either set the engine at 1000 to 100,000 nodes per move or use centipawns to find brilliant moves. I guess it has to be a brilliant move when the move is rated higher than it is humanly possible. https://en.wikipedia.org/wiki/Instructions_per_second https://lczero.org/dev/wiki/lc0-options/ https://www.melonimarco.it/en/2021/03/08/stockfish-and-lc0-test-at-different-number-of-nodes/ https://www.gpucheck.com/gpu-benchmark-graphics-card-comparison-chart https://www.cpuagent.com/ https://www.phoronix.com/news/LCZero-NVIDIA-Benchmarks https://www.phoronix.com/review/nvidia-rtx3080-compute/6 https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units https://www.cgdirector.com/best-computer-3d-modeling-rendering/ https://medium.com/the-mission/how-to-build-the-perfect-deep-learning-computer-and-save-thousands-of-dollars-9ec3b2eb4ce2 https://lucaschess.pythonanywhere.com/static/pdf/english/Lucas%20Chess%204-Tools.pdf FEN (Forsyth–Edwards Notation) GUI (Graphical User Interface) PUCT (Predictor + Upper Confidence Bound tree search) algorithm. SAN (Standard Algebraic Notation) UCT (Upper Confidence Bound for Trees) Has anyone else tried to find brilliant moves in popular opening games?