@Toadofsky said in #25:
>
I may have misunderstood the revealing. Actually I hesitate, about the quote being first degree or second degree. As in certain contexts, I do like to think while assigning some black-box attribute to parts of a model, often depending on the UI context of visible input- and visible output. But here.. we have the source-code, some recently improving body of documentation that is not just for internal dev. audience anymore, and the slowly evolving UCI command line interface that surely lichess has access to.
Some of your view in the past, seem to have been about prudence in sharing the "instability" at depths of SF search trees for user input position. I do not agree to that, because, it strenght that "trust" by omission of information.. That is one thing.
but that is arguably research mindset. What I have been trying to be more suble about is the extra black box that is unnecessary between user and lichess layer about UCI-SF. I am not talking about full UCI output even.. (although it could help users helping lichess, as even if dev-illiterate, chess players are a bunch of rational grinders with some imagination.. that can at least sound board things for lichess. or open mind doors... when not the annoying user side complainer, that some image of them might carry, me not having that image, but I sense overworked human resources might).
I am talking in the direction of the blog trying to find more information helpful for the user in learning mode.. PEDAGOGY!
In making more than one UCI parameter regime (ultimate depth, more and forever....), to get more than one "metric" (metric has a math. meaning for me) or measurement dimension, or evaluation scale for the same position evaluation (in context of game sequence).
Objective is more information about the nature of the user human chess from SF abilities to discover best moves but also some ordering of lesser moves..
research might be needed to establish an automatic way, to not need the user to use the full PV output (that lichess is adamant to spare the user some damaging exposure to), I have surrendered taoistically to that state of reality.
So. the knowers behind the UI of lichess, might just have to use shallower input depth or just follow the trace of the iterative deepening for the depths are which the called out full depth score difference "blunder" class of user move choice, at which SF first detects such difference from its leaf evaluations... I think I would be glad to help finding such things as preliminary reasearch to help the developpers. But working against cultural inertia is not a great motivator. Actually I suspect the devs would be able to figure it out. but if there is not intent to improve the use of the engine because of the expanding black box boundary spare the credulity of the user base mentality, such efforst become impossible to even imagine.
I have already seen some lichess user colleague (also chess mentor of mine) do a manual approximation and discuss it with me, using both lichess and command line.. (to get the reall thing, lichess forcing some ply per ply jumping added perturbations)
from shallow leaf evaluatoin spectrum (gamut or whatever, look at few top PVs wide enough mayeb) increase input depth (or follow iterative or the TT inputs before memory loss), until the ultimate depth previously called error shows its face...
then one could start leanring about things like was it a positional blunder (long term to conversoin or visible effect in the face at position) and short term blunder.. as we, learning users, (hence PEDAGOGY word before), do not do ultimate depth errors all the time.. In order words "depth to discovery" from the point of view of SF during iterative deepening. BTW if instead of seconds in the UI we could have max input depth control, together with PV width, we could as user figure it out.. but if lichess insists to spare our credulity and trust in the black box beyond UCI, then what i just suggested. automatize shallowest error characterization for enhance chess insight feedback to users.
some of us aspire to more than just oracle views of our chess play. some of us would like to have long arc story understanding of our own play.. some of us, are able to improve tactics elswhere than in our full games.
I just think it is a continuation of the full game contextualization or error feedback that lichess has started in adapting the SF output in its 3 bin error classicatoin for pedagogical use of SF.
one might have to accept that there is a pedagogical value in making user aware of the depth dimension of their errors.
that maybe game are not just strings of tactical segments patch à la queue leu leu... until last blunder ....
I would be glad to help.. this is inspireing. And is related to the blog, no joke. The blog is walking in that issue of depth to conversion.. and handling of imblanace over chess time.. having more than one UCI-SF point of view.. is a begginning.
not sure that just using the one D score along the chess time axis can give us that, as they blog seemed to have attempted.. or i did not understand yet.