- Blind mode tutorial
lichess.org
Donate

Finding Threats and Piece Manoeuvres with Engines

#2

If one could force the engine to play "human" moves for the other side, the ideas might show up more clearly. But classifying moves as human is even more difficult.

What is a "human" move?

> If one could force the engine to play "human" moves for the other side, the ideas might show up more clearly. But classifying moves as human is even more difficult. What is a "human" move?
#3

@Toadofsky said in #2:

What is a "human" move?
I meant moves that sacrifice material just to prevent a strong plan. For example, in Short-Timman Stockfish wanted to give up the queen after White played Kg3. It's objectively better than not doing it, but humans usually don't give up so much material when they don't even get any counter play for it.
There are many other examples where engines give up material in bad positions without getting much counter play because they see that far down the line they can't keep it anyway. GM Matthew Sadler recently published a video where he showed a game where LC0 didn't take a pawn even though it wouldn't have given Black much counter play, preferring instead to focus on a bad bishop: https://www.youtube.com/watch?v=4UzEOOI3Obs

@Toadofsky said in #2: > What is a "human" move? I meant moves that sacrifice material just to prevent a strong plan. For example, in Short-Timman Stockfish wanted to give up the queen after White played Kg3. It's objectively better than not doing it, but humans usually don't give up so much material when they don't even get any counter play for it. There are many other examples where engines give up material in bad positions without getting much counter play because they see that far down the line they can't keep it anyway. GM Matthew Sadler recently published a video where he showed a game where LC0 didn't take a pawn even though it wouldn't have given Black much counter play, preferring instead to focus on a bad bishop: https://www.youtube.com/watch?v=4UzEOOI3Obs
#4

I have a TODO item to make this happen, too. One of the ideas that I had is to run SF with multiple PVs and then group the lines on winning and losing, then try to find common patterns. You would get stuff like thematic moves where White wins if they get their pawn to c4, but Black wins if they get their bishop somewhere, regardless when that happens. Another pattern would be getting rid of a certain enemy piece, no matter where on the board.

I have a TODO item to make this happen, too. One of the ideas that I had is to run SF with multiple PVs and then group the lines on winning and losing, then try to find common patterns. You would get stuff like thematic moves where White wins if they get their pawn to c4, but Black wins if they get their bishop somewhere, regardless when that happens. Another pattern would be getting rid of a certain enemy piece, no matter where on the board.
#5

You can try playing 'nothing moves', essentially passing the move over. For example in the Short game after 1. Kh2! play Ba8-Bb7 and see what the comp does.

You can try playing 'nothing moves', essentially passing the move over. For example in the Short game after 1. Kh2! play Ba8-Bb7 and see what the comp does.
#6

Finding threat featre is like a null-move pass your turn thing. If there one passes one's turn here might be what the one sided oppononent threat spectruym might be (might be nice to have PVs too, for the "idea" is rarely alone.

moved to musing the rest. I just leave the suggestion here.

Finding threat featre is like a null-move pass your turn thing. If there one passes one's turn here might be what the one sided oppononent threat spectruym might be (might be nice to have PVs too, for the "idea" is rarely alone. moved to musing the rest. I just leave the suggestion here.
#7

Manoeuvres

is that another word for combinations. or if only one piece, its multi turn potential paths abstracting the current obstacles, or assuming some persitent or location stablility. Did you not have somethign to say about that in some previous blog. I am glad you are reviewing all your ideas. one small chunk at a time. (well, and also the question/reasoning process, I value sharing the nature of the question, and its openness a lot, less passive reading).

I am reading ahead perhaps.. i read better that way. either wrong or right. there is a micro-debate stake for the reading. got to fool my own "animal" or what do other say, the "it" in archaic terminology? or the chat bot in me... is marketting terminology? that was out of control, sorry.

answering self. I must have seen the very first diagram, and it came out as if I was guessing... ahh, the subconcsious tortuous paths....

> Manoeuvres is that another word for combinations. or if only one piece, its multi turn potential paths abstracting the current obstacles, or assuming some persitent or location stablility. Did you not have somethign to say about that in some previous blog. I am glad you are reviewing all your ideas. one small chunk at a time. (well, and also the question/reasoning process, I value sharing the nature of the question, and its openness a lot, less passive reading). I am reading ahead perhaps.. i read better that way. either wrong or right. there is a micro-debate stake for the reading. got to fool my own "animal" or what do other say, the "it" in archaic terminology? or the chat bot in me... is marketting terminology? that was out of control, sorry. answering self. I must have seen the very first diagram, and it came out as if I was guessing... ahh, the subconcsious tortuous paths....
#8

Quiet positions are by definition absent of threats, so looking for threats won't explain the moves played

Are you sure the engine is not just showing the best move according to switch turn game. The word "threat" is not only about capture threats. I understand you mean tactically not hurting immediately. And that the human detectable might be more tangible near a desirable capture, as basicaly all satisfacotry puzzles that don,t have mate solving last solutoin step should show the person who just got the credit for solving the puzzle. That the positional non capture based analysable signal are difficult to use as objective tool that could inform any patzer level of learner.

But my previous rambling** arugment that ideas hide other sub-ideas that might all still influence the dynamics or the thinking many turns ahead, as possible alternative in the real chess context on not best play (strong chess problem, as strongly solving chess might be about solving all continuations of all legal positions, no matter how not best play obtained, the feared non cooperating opponent when having studied hard on some known to be safe mainline, in vogue. well that argument might be
that earlier threts or deeper branches outside of the best PV lichess truncate line but including it, might the wider the PV cross section, the more hidden threat in such entrails (I can,t see that, please put that obscene things away from my sensitive eyesight).

I am some chess ranting catching up to do. taking a vacay from chat bot immersion (and struggle to thin the fog behind the claims, as a mere user that got baited by the offer). so pardon the from nowhere, intensity. it comes from there. Got some steam accumulated from there. So I try to be constructive about it, but then, I might let some tone slip out.. thanks for bearing with me.

I think i also possible just suggested something about preceding question. Problem reamins, if we had easy access without contortions or command line tyranny (some steam left i guess!) , to the entrials of the single position moderate search tree, or the ability to query it, having some structuring distributions of node types, at least all the evaluated ones, could serve as some signature of the tree, or those leaf not quiet enough to justfiy extersions, those leafs not controlled by user input parameters if they could have some depth control (so not on Lichess yet, and i bet not many online platforms that integrate many chess tools to make a chess home for chess lovers).

The problem would remain of which deeper nodes to consdier. But my width suggestion might be countering your point that the best line would not have considered as premptively blocking it. but that information would only be visible in the sup-optimal play.

I now think we might the same neigborhood of concerns and realization about how we are wasting information sticking to oracle chess theory of learning. I like evocative punch lines. That might be caricature, but not entirely wrong.

I.e. we want entrails too....

** (the rambling moved from this discussion, sometimes i get some self conscious lucidity flashes)

> Quiet positions are by definition absent of threats, so looking for threats won't explain the moves played Are you sure the engine is not just showing the best move according to switch turn game. The word "threat" is not only about capture threats. I understand you mean tactically not hurting immediately. And that the human detectable might be more tangible near a desirable capture, as basicaly all satisfacotry puzzles that don,t have mate solving last solutoin step should show the person who just got the credit for solving the puzzle. That the positional non capture based analysable signal are difficult to use as objective tool that could inform any patzer level of learner. But my previous rambling** arugment that ideas hide other sub-ideas that might all still influence the dynamics or the thinking many turns ahead, as possible alternative in the real chess context on not best play (strong chess problem, as strongly solving chess might be about solving all continuations of all legal positions, no matter how not best play obtained, the feared non cooperating opponent when having studied hard on some known to be safe mainline, in vogue. well that argument might be that earlier threts or deeper branches outside of the best PV lichess truncate line but including it, might the wider the PV cross section, the more hidden threat in such entrails (I can,t see that, please put that obscene things away from my sensitive eyesight). I am some chess ranting catching up to do. taking a vacay from chat bot immersion (and struggle to thin the fog behind the claims, as a mere user that got baited by the offer). so pardon the from nowhere, intensity. it comes from there. Got some steam accumulated from there. So I try to be constructive about it, but then, I might let some tone slip out.. thanks for bearing with me. I think i also possible just suggested something about preceding question. Problem reamins, if we had easy access without contortions or command line tyranny (some steam left i guess!) , to the entrials of the single position moderate search tree, or the ability to query it, having some structuring distributions of node types, at least all the evaluated ones, could serve as some signature of the tree, or those leaf not quiet enough to justfiy extersions, those leafs not controlled by user input parameters if they could have some depth control (so not on Lichess yet, and i bet not many online platforms that integrate many chess tools to make a chess home for chess lovers). The problem would remain of which deeper nodes to consdier. But my width suggestion might be countering your point that the best line would not have considered as premptively blocking it. but that information would only be visible in the sup-optimal play. I now think we might the same neigborhood of concerns and realization about how we are wasting information sticking to oracle chess theory of learning. I like evocative punch lines. That might be caricature, but not entirely wrong. I.e. we want entrails too.... ** (the rambling moved from this discussion, sometimes i get some self conscious lucidity flashes)
#9

To find manoeuvres, I looked if the best engine line consists of moving one piece multiple times in a row. Then one can look for the square where the piece will eventually end up on, instead of focusing on the individual moves.

I assume you mean that this can be made automatic. Could you not have a better fishing crop if relaxing or extending your search to also suboptimal branches? while having the best line.

or are you walking us to the ideas, that might be worth devising automatic search functions with the full PV as input (and my suggestion as above. not look at the bigger chess than just that line, your argument about the engine already having done all the suboptimal weeding out when just looking at best line is likely to the human reachable ideas worth considering within human horizon (breadth) or in non turn by turn derisable imagination long shots thinking (yes, we do that, we humans, or I do, and I am loving it, right or wrong that is the beauty of chess for me, with the win convened goals as ruleset, like any other rule, the official goal is winning, and evaluation of odds given position is the real many game life long goal). enough poetry. What do you think about looking not only at best line, applying your arguemnt that the engine might have had all the remote future resrouces to have hidden all of those insightful threats..

> To find manoeuvres, I looked if the best engine line consists of moving one piece multiple times in a row. Then one can look for the square where the piece will eventually end up on, instead of focusing on the individual moves. I assume you mean that this can be made automatic. Could you not have a better fishing crop if relaxing or extending your search to also suboptimal branches? while having the best line. or are you walking us to the ideas, that might be worth devising automatic search functions with the full PV as input (and my suggestion as above. not look at the bigger chess than just that line, your argument about the engine already having done all the suboptimal weeding out when just looking at best line is likely to the human reachable ideas worth considering within human horizon (breadth) or in non turn by turn derisable imagination long shots thinking (yes, we do that, we humans, or I do, and I am loving it, right or wrong that is the beauty of chess for me, with the win convened goals as ruleset, like any other rule, the official goal is winning, and evaluation of odds given position is the real many game life long goal). enough poetry. What do you think about looking not only at best line, applying your arguemnt that the engine might have had all the remote future resrouces to have hidden all of those insightful threats..
#10

The manoeuvre is the winning idea, but Stockfish prevents it from happening.

So that was human chess associate to the human knoweldge from the repertoir example used. As you were having a control example with current consensys main idea this was discoverd as new opening branch one and still maintained as "winning". The winning being about population of high level games in some past having made it clear to all (or enough that would keep propagating the good news as knowledge). While the engine executable instance is having its own programmed in version of chess knowledge, that is not in this case directly based on outcome, but explicit theory of evaluation evolved through many versions, and approximate nowadays suing NN netowrds to spread the evaluation model upstrean by generalization model fitting (the simple eval being the programmed in model of evalution, that de facto wins in established engine game format pools with all compteting engines).

The winjing idea is from human chess culture. And I value that more than the engine winning culture, as not well defined and sorry to say, not of the same "genetic and phenotypic" diversity as I would expect the history of human chess to have maintaind by the mere small brain principle that even if we would try to copy each other deeply, we would not be able to do reproducibly. We are good that way. Human error to the rescue of human chess.. .. i stop.

My idea is already clear. Get over the assumption that winning is so well defiend. in either world. And use the blurry visionh method.. to find what stands out.

> The manoeuvre is the winning idea, but Stockfish prevents it from happening. So that was human chess associate to the human knoweldge from the repertoir example used. As you were having a control example with current consensys main idea this was discoverd as new opening branch one and still maintained as "winning". The winning being about population of high level games in some past having made it clear to all (or enough that would keep propagating the good news as knowledge). While the engine executable instance is having its own programmed in version of chess knowledge, that is not in this case directly based on outcome, but explicit theory of evaluation evolved through many versions, and approximate nowadays suing NN netowrds to spread the evaluation model upstrean by generalization model fitting (the simple eval being the programmed in model of evalution, that de facto wins in established engine game format pools with all compteting engines). The winjing idea is from human chess culture. And I value that more than the engine winning culture, as not well defined and sorry to say, not of the same "genetic and phenotypic" diversity as I would expect the history of human chess to have maintaind by the mere small brain principle that even if we would try to copy each other deeply, we would not be able to do reproducibly. We are good that way. Human error to the rescue of human chess.. .. i stop. My idea is already clear. Get over the assumption that winning is so well defiend. in either world. And use the blurry visionh method.. to find what stands out.