This follows another questions here: What is the status for EIGHT piece endgame tablebases?

My answer there was appropriately removed as being question in reaction to that question, well in light of the question. So here is that non-answer, modified to be a question. I wish it to be considered with the above link in mind, as this previous question follows the existing expansion from 3 men to 7 men. As if it was the only way to use such information database scheme (the work is more than the tables we use).

Is there some machine learning approach that would use existing TBs and dissect their nesting statistically properties for clues of "nested" patterns?

I find that we might be waiting for computer power to do the exact retrograde, why not looking for more human ways of finding retro gradable, human theories of chess, that might be corroborated by information contained in some subsets of the TBs.

I find that, as humans, we actually started learning chess, each of us, on some naked board, one piece at a time. Why some jump to 32 right away? I forgot. But now I am seeing endgames, and compositions around them (TB is not just legal chess positions from standard IC), as research or study challenge that can stretch our reasoning to understand the un-stretched.

** So not just practical endgames, but games that allow both rational theory building from the spatial logic likely contained in the human created finite ruleset of chess, and also our experience. With an intrinsic pool of player independent referential "basis" of data, pure chess information. Not a statistical pairing outcome extrinsic accuracy measure. We have microscopic ruler. We don't have to follow it blindly, but we have that, being exact, we can approximate from there, do all sorts of data analysis or ML feature reduction transformations to test logical reasoning hypotheses or even suggest relations to investigate rationally, that can be controlled and characterized back to that level.

I find that from a chess science (not data hoarding), point of view, there might be untapped human accessible clusters of information from already existing from 3 chessmen up.. 3 to 7, that is four nesting levels.

I also know that there have been attempts to compress with Lc0 NN, up to the 6 men EGTB, as is (i.e. full TB, even that which might not be legal, just saying there is that chess, and there is human chess, not all about combinatorial turn by turn retrograde, and that is still logical, using spatial abstractions that consider that not every distinct position is equally distant w.r.t. to their solutions construction, itself not only accessible from turn-by-turn lowest logic. This might be a different kind of chess complexity, not having any chess words going for it, and not much computer terminology either, which is based on every distinct position being distinct equally. (not having the embedding to see any other angle).

I can improve on the question, upon feedback. The point here is about extracting more than oracle value from EGTB chess board analytical tool. As we have visuo-spatial abilities that are not put to work. Individually, there have been specific problems of chess or endgame classes that might have work on extracting such information, but it has not been very methodological as far as I could be made aware so far. Perhaps here, we could gather such attempts, if my top question, is a bit premature.

  • Scientific agreement seems to be that the pattern needed in the understanding of RR/RN etc. are completely ineffable to the human mind. Neither would I use a TB of KR/K in training of mating with a lone R - it's faster but a bit unsystematic. Commented Nov 24, 2023 at 9:18
  • I agree that the knight mobility figure, while still bound to combine with existing king board graph (8x8) and its 8 spatial lines (kind walk directions, and adjacency), is not a single ply slider type of mover (its figure is not simply connected like other pieces). Making for a more combinatorically heavy task to calculate. However it is still bound to notions of distances on the multi-move scale. It mid-range board covering that allows it to jump over fences, is still leaky. Alone it needs help to "corner" a king. I agree that TB as oracle is not very good for human reasoning.
    – dbdb
    Commented Nov 24, 2023 at 21:40
  • So my question has w aspect, one being to gather the background about the scientific agreement about the classes of endgame position input (I like seed) or problem that the chess multi-move solutions provided by using TB data as solo thinker and local analysis. TB output only considers the full branching of legal moves from input position, the continuation being perfect play given the termination ruleset beyond mate, stale mate and fold repeat, i.e. the 50 or 70 sanity limit for tournaments.
    – dbdb
    Commented Nov 24, 2023 at 21:46
  • So I am not interested in oracle turn-by-turn reasoning. TB does not only give that. I know of one chessable course (i did not study it yet), where it is actually used to create category of placement arrangements (patterns) that provide for planning level of reasoning of a human visuo-spatial type (positional), that the turn-by-turn would have to follow anyway. This is not about using TB as turn by turn oracle trainer. But as you point out mixing an endgame short range piece with a very teleporting piece can lead to lots of wandering. That I have explored, partly with TB for its abyss.
    – dbdb
    Commented Nov 24, 2023 at 21:59
  • The point of my question is to further such plan level problem solving with the help of the TB not from the single player thinking challenge but for the information as database of all possible games. We have not only one game tree, but all possible continuations for all legal and non-legal seed positions contained in the input space from retrograde analysis by machine symbolic computation (guessing). There are full data set methods to extract pattern hypotheses that we can't uncover alone turn by turn as exhaustion limits us. Good examples though, on the challenge side.
    – dbdb
    Commented Nov 24, 2023 at 22:04


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.