Engines lines are sometimes very crazy. Like, spmetimes, even after 1h of checking engine lines, I still cant understand the idea behing the move.
not good!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Hmm... Useful information for beginners and advanced players. I advise you to play with different levels of 3+0 stockfish and you will see how well you play chess. I checked - it helps!
the same happens in real life!
I am an accountant and studied accounting in the university
when I was in university so many things were looking unclear but after working accounting practically I found out none of them were hard!
the same in programming! these days I am learning python but I only read and not trying to write a code!
that is why I think programming is tough but in reality it is not!
the best way to learn something is by trying and failing and then you can learn! you can't learn if someone tells you should do that kind of thing. you need to try it by yourself and using other people's experience from time to time!
I am an accountant and studied accounting in the university
when I was in university so many things were looking unclear but after working accounting practically I found out none of them were hard!
the same in programming! these days I am learning python but I only read and not trying to write a code!
that is why I think programming is tough but in reality it is not!
the best way to learn something is by trying and failing and then you can learn! you can't learn if someone tells you should do that kind of thing. you need to try it by yourself and using other people's experience from time to time!
any word about how to trust the engine scores at current position with respect to the depth of search, and displayed PV depth at 16 or less? often not the input depth.
Or deep search "instability" beyond certain depth. I think it is not about the score so much as to the PV being output by engine.
Lichess caps this sharing of the engine analysis, for such reason of the PV beyond 16 plies not being trustworthy. It has been confirmed in one issue on github from a concerned dev of SF. not documented but it made me accept lichess prudence.
Problem is that the mysterious score at current position, is propagated back from the very deepest position of such deep tree search. So I find that a waste of information, even if the info has its problem, the score is still associated to such position information. And having such position would allow own human fully aware of "instability" of said propose PV, to do own current position score to position feature at that untrustworth depth (but still explaining the score in ways where the human can actally make an opinion on the score given the visible static feature at depth).
I am convoluting and being overly precise above.. because I find that "Trusting your engine" blindly is asking for tunnel vision inertia or absence of ways to use the engine feeback with full human abilities at hand. There is learning about chess that is being ignored.
And if more humans were to access such information, maybe there would be pressure for more analytically accurate or with confidence for static leaf evaluation function in the future. winning a long game can happen on very few moves along the game... it does not garantee that each position was being internally analysed best, but good enough and knowing how coding was being amended, fast enough. With more human awareness the instability might become part of the human analysis concerns.. There might be instablity, but why not allow the human using the engine as analysis tool, to judge what is not instable, that current position score and the deep static evaluation of the leaf (be it from an "unstable" beyond 16 ply tail).
thanks for you blog. It gave me the opportunity to rephrase something that has been bothering me for a while.. each iteration brings a better more on topic version..
Or deep search "instability" beyond certain depth. I think it is not about the score so much as to the PV being output by engine.
Lichess caps this sharing of the engine analysis, for such reason of the PV beyond 16 plies not being trustworthy. It has been confirmed in one issue on github from a concerned dev of SF. not documented but it made me accept lichess prudence.
Problem is that the mysterious score at current position, is propagated back from the very deepest position of such deep tree search. So I find that a waste of information, even if the info has its problem, the score is still associated to such position information. And having such position would allow own human fully aware of "instability" of said propose PV, to do own current position score to position feature at that untrustworth depth (but still explaining the score in ways where the human can actally make an opinion on the score given the visible static feature at depth).
I am convoluting and being overly precise above.. because I find that "Trusting your engine" blindly is asking for tunnel vision inertia or absence of ways to use the engine feeback with full human abilities at hand. There is learning about chess that is being ignored.
And if more humans were to access such information, maybe there would be pressure for more analytically accurate or with confidence for static leaf evaluation function in the future. winning a long game can happen on very few moves along the game... it does not garantee that each position was being internally analysed best, but good enough and knowing how coding was being amended, fast enough. With more human awareness the instability might become part of the human analysis concerns.. There might be instablity, but why not allow the human using the engine as analysis tool, to judge what is not instable, that current position score and the deep static evaluation of the leaf (be it from an "unstable" beyond 16 ply tail).
thanks for you blog. It gave me the opportunity to rephrase something that has been bothering me for a while.. each iteration brings a better more on topic version..
@Bergie2000 said in #2:
> Engines lines are sometimes very crazy. Like, spmetimes, even after 1h of checking engine lines, I still cant understand the idea behing the move.
try to read my post. it is related. That is for the remote conversion detection feedback that you will never see on lichess PV display. as you increase depth the PV stays withing its caps. so one feels the need to manually update the current position, by dumping some node of the early PV nodes (safest against horizon problems is the leftmost). But then you keep also pushing the PV leaf further, still not shown on the engine best PV line display.
There is another problem about using an engine bred for tournament ELO measures. It gains ELO by cutting corners. On width. that is long story. And could be argued many ways. but all people who would discuss would agree it has been the trend and explaination of high ELOs (and the engine tournament spec historical evolution if any).
My understanding is that ELO gain can and has been had by favoring speed over width certainty (or pure AB design, itself relying on leaf degree of evaluation relevance (accuracy? certainty?) to winning outcome probability within its continuation sub-tree). The belief that depth will solve all past errors, may be conforting that such compromise has no accuracy consequence, because anyway , people are satistifed by knowing that it beats anything with many games and ratings.. That ought to back propagate to single position engine scoring..
I say this, because the time saving heuristics and code optimisation hard deeply (another kind of deep) intertwined and hard wired in the main search (and its children) algorithm. Giving it more time, is just increasing depth, in the context I claim above.
The only user controllable analytic tool setting that could affect that departure from pur AB is forcing mutliPV (lichess has 5 max, but SF has 20 I think), where it might be slower overall, but you know you are pushing against SF tendency to think that narrow and deeper is always better than shallow and wider.
However i think your post is about interpretability of the engine current position score value, and its output proposed main PV (which is never visible for your human eyes.. like you can't handle that kind of departure from blind trust).
But "Trust you engine"!. Like someone wanting something out of you asking you "Trust me!".
> Engines lines are sometimes very crazy. Like, spmetimes, even after 1h of checking engine lines, I still cant understand the idea behing the move.
try to read my post. it is related. That is for the remote conversion detection feedback that you will never see on lichess PV display. as you increase depth the PV stays withing its caps. so one feels the need to manually update the current position, by dumping some node of the early PV nodes (safest against horizon problems is the leftmost). But then you keep also pushing the PV leaf further, still not shown on the engine best PV line display.
There is another problem about using an engine bred for tournament ELO measures. It gains ELO by cutting corners. On width. that is long story. And could be argued many ways. but all people who would discuss would agree it has been the trend and explaination of high ELOs (and the engine tournament spec historical evolution if any).
My understanding is that ELO gain can and has been had by favoring speed over width certainty (or pure AB design, itself relying on leaf degree of evaluation relevance (accuracy? certainty?) to winning outcome probability within its continuation sub-tree). The belief that depth will solve all past errors, may be conforting that such compromise has no accuracy consequence, because anyway , people are satistifed by knowing that it beats anything with many games and ratings.. That ought to back propagate to single position engine scoring..
I say this, because the time saving heuristics and code optimisation hard deeply (another kind of deep) intertwined and hard wired in the main search (and its children) algorithm. Giving it more time, is just increasing depth, in the context I claim above.
The only user controllable analytic tool setting that could affect that departure from pur AB is forcing mutliPV (lichess has 5 max, but SF has 20 I think), where it might be slower overall, but you know you are pushing against SF tendency to think that narrow and deeper is always better than shallow and wider.
However i think your post is about interpretability of the engine current position score value, and its output proposed main PV (which is never visible for your human eyes.. like you can't handle that kind of departure from blind trust).
But "Trust you engine"!. Like someone wanting something out of you asking you "Trust me!".
In my opinion, chess engines should be only for checking your own opinion
they can help often after enough banging one's head on a problem. but not as master of your own thoughts.
@ultimatemachine yeah, i completely agree with you that's really the best way to learn. especially after OTB games, do an exhaustive analysis on your own first (even better is to analyze with your opponent, and get their thoughts about critical moments too), and only then use the engine to blunder check later. the engine will almost always point out a resource that you and your opponent either missed or rejected and then it's up to you to decide whether the engine line is something you can learn from, or whether the line is just too complex or inhuman