after thought, about picking apart engines to combine parts that can fit with each other.. (which is the exact off-topicness of the post i was going to delete).
Crafty comes to mind..
It did have that possibility in mind. I have no experience, just read some of its documentation, one of the rare case of interpret-able engine documentation, but it was in the concept of its design I think: to be a engine development teaching exemplar.
It had some sort of "user" control of which heuristic would be made live in some instances of its family** of engines.
It may be a user that can tweak, and do some parameter changes... ok.. I may also have it wrong about crafty.. But i just thought it would be an example.. of something usable and already designed to keep some modularity with outside handles.. (trying to explain... use salt).
** I could not help but inject some notion of engines not being instances until all its parameters are given values, but families of engine can exist with symbol parameters.
Looking at crafty, if I remember right about its modularity with "user" handles almost at input, it might be viewed as a family of engine or the functions the parameters would span, as set of functions (this is one kind of mathematical view point, figure out the inputs and outputs to make it work).
https://www.semanticscholar.org/paper/Imitation-Learning-by-Estimating-Expertise-of-Beliaev-Shih/1467cffb0bad7163412b03ab473f44a54c5c40ac
https://api.semanticscholar.org/CorpusID:246485360
Corpus ID: 246485360
Imitation Learning by Estimating Expertise of Demonstrators
M. Beliaev, Andy Shih, +2 authors Ramtin Pedarsani
Published in ICML 2 February 2022 Computer Science
It appears i may have use my own words.. after some time, that is how i might remember things i digested. But this is the paper that seems compatible with notions of biases and complementarity of them.. The notion of expertise regions.
This might be related and closer to chess.. (Style not far from bias). Nothing says that the bais is wrong. One may learn to adjust own bias.. The region of expertise is likely related to the individual experience trajectory (which is likely diverse with humans).
And guess what, if all the human tournaments were making sure that many wildtype biases were still around through the funnellihng all the way to the top, it might very well be argument for the top, to be the least bias of humans....or the best most uniform one.... But we might not know until we ask a few questions or accept to consider them.
DOI:10.48550/arXiv.2208.01366Corpus ID: 244427643
Detecting Individual Decision-Making Style: Exploring Behavioral Stylometry in Chess
Reid McIlroy-Young, Russell Wang
Published 2 August 2022 Computer Science ArXiv
https://api.semanticscholar.org/CorpusID:244427643
The last posts. are the end of this tangent on this thread from me.. i hope I am not exaggerating.. Thanks to op for allowing me such excursion.
Disclaimer: I only read the abstracts.. (may have dove into some chunks too.. don't remember).. but if those have some echo here. it might be motivation for me to dig into them further.)
TCEC (Top Chess Engine Championship)
2022 live Final is in full swing now on TCEC-CHESS
Stockfish vs LC0
Best in one phase of the game does not mean they are the best engines. Just need to be best in one phase of the game.
Best in one type of opening is even more useful.
@EvilChess said in #18:
>
Yes, and actually the score attributed to a single root search best move, is the score attributed to the leftmost node of the first PV.
I assume with different vocabulary, same for leela candidates. It is not however doing the same tree exploration and full position information evaluation selection upon its search breadth (that is as far as i venture to say, i still am in need of understanding about how the search is stopped and the statistics involved). My current understanding is that it uses confidence level as stopping criterion.
question to you @EvilChess. about my current fog about leelas.
MCTS, PUCTS, ? family of tree search, or is it graphs... as my understanding of the math. behind A0 was that of a position space and an action space, not of decision trees directly. Although such tree can be considered describing different paths of successive actions on the position space** or set, as each path being the result of the decision to not follow the other paths in a give game instance.
** space would need some kind of structure, but hey, why not jump there and see where that leads, in my mind even a set might have a 2D layout a priori, until more information requires a third D, and then some fog, Are you able to figure out where I am lost?
@Toscani said in #25:
>
you would look at performance ELO.. perhaps. but how many games with those constraints.. (i was thinking of the repertoire "best").
for phase i am not sure how even ELO might help. What does it mean to be better in one phase than the other engine. that it loses advantage according to human or another engine... I think we might see bad endgames more easily, than bad middle-game... unless these are tactical blunders in our face (those we can see). but if long term mistakes, and not visible to us, as only remote visible loss, i don't see that we have the information to tell.. Or my eyes are not both in their sockets at this early time of day. Which is possible.
And I would like to remind that in context of just ELO in such anemic pools, what we get is "better than" not best.. (just a connotation difference, keeping in mind the nature of the measure being used). It could also be argued: "faster than".
@dboing said in #26:
> ** space would need some kind of structure, but hey, why not jump there and see where that leads, in my mind even a set might have a 2D layout a priori, until more information requires a third D, and then some fog, Are you able to figure out where I am lost?
Hmm no that isn't clear. If it is about possibility of an engine to project some possible board possible future positions, based on partial board analysis, and verify their feasibility, then it would be related to an old thread of mine, but not this one.
@EvilChess said in #28:
> Hmm no that isn't clear. If it is about possibility of an engine to project some possible board possible future positions, based on partial board analysis, and verify their feasibility, then it would be related to an old thread of mine, but not this one.
I will get back to you.. right not here. off topic. (it was just in case it was obvious to you, but i will try to find your thread anyway).
Edit: https://lichess.org/forum/general-chess-discussion/books-on-the-mathematical-side-of-chess-?page=1
how could have I missed that one.. That is the problem with forum having been demoted from the lobby. Easy to miss interesting threads now. (at least faster loss of thread titles awareness, since we only have our time line not the forum general time line, in lobby). I will look at it.. But my question was about representations of chess games and how did trees as in MCTS came about from what appeared to me as more of a graph representation with unique positions upon which "actions" were applied as directed edges (defined as I understood then, as transitions between those positions). However, I think that the policy head has moves of material units as actions not FEN to FEN pairs (if FEN were how the unique positions were implemented, in reality position is being exploded in such a way that the moves are just edges or notches in that representation at input layer, i might have lots of misconceptions...). still off topic though..for another time too.. letting you know where i am been stuck for a long time.
@dboing said in #29:
> But my question was about representations of chess games and how did trees as in MCTS came about from what appeared to me as more of a graph representation with unique positions upon which "actions" were applied as directed edges (defined as I understood then, as transitions between those positions). However, I think that the policy head has moves of material units as actions not FEN to FEN pairs (if FEN were how the unique positions were implemented, in reality position is being exploded in such a way that the moves are just edges or notches in that representation at input layer, i might have lots of misconceptions...).
Whatever representation used, if it is complete, then it is interchangeable with FEN. For any given way of representing the board, if any instance of it fully identifies single board position, and it also carries info that indicates about all of the additional features that a FEN notation can specify (such as castling availability), then it convertible to FEN. And if that representation can represent all possible board positions, then any valid FEN notation can be converted to this representation as well.
Why does it matter which representation to use then? Because each one allows a different set of analysis and inferences to be applied. We cant derive much from a FEN notation, but from conventional board representation stored as a few bit matrix layers, it's easier to check some stuff, such as which pieces are under attack.
And it happens that Leela uses a matrix representation, but that isn't the best chess representation for a NN engine. Because, the matrix offers poor/limited possibility for convolutions. In Leela, the convolutions take 3x3 parts of the matrix. That does make it stronger, but not so much, because a 3x3 board slice doesn't have much meaning. With a graphs representation of the board, however, we can properly define the convolutions, and that should be the "game changer" feature of chess NN engine.
Ok, I hope my comments help. I'm done now with this topic by now, because it can take time and I don't want to get even shorter in time than I'm now.
This topic has been archived and can no longer be replied to.