3

There's an ongoing trend towards building 'Explainable' AI so that not just the results of a decision process are generated, but also some form of human-readable explanation for it. The Wikipedia article mentions that some work has been done on building explanations for decision trees, and I was able to find DecodeChess (link not given since it's a commercial entity, but it's surely not hard to find) which purports to be using XAI for positional analysis, but a little digging hasn't turned up any mention of the ideas being turned towards endgame tables in hopes of finding something more like an algorithm for the complicated endgames (e.g. KRB vs. KNN) where many of the moves seem almost entirely random. Is anyone aware of research done in this direction?

3
  • 1
    Would the downvoter care to explain? I'm happy to revise this question as appropriate. Dec 2, 2021 at 18:15
  • 3
    (I didn't downvote by any means), explainable AI is generally an idea that is restricted to (typically blackbox) AI/machine learning. Endgame tablebases are certainly not machine learning and not even really AI, but instead gametree mappings. So it's not possible to 'explain' the internal logic as it is for Stockfish (which has a positional evaluation based partly on many detailed heuristics), because there is no internal logic (except 'whatever the subsequent moves, this works'). Dec 3, 2021 at 3:13
  • Black box, boundaries are up to us to decide. The data is right there, inside the trained machine. One has to ask the right questions, or even ask them first. Not agreeing with SF being explainable on the basis of its leaf evaluation function having had source code with algebraic functions of position information that had names as "explanation". Those have parameters with values to optimize and a whole spaghetti of algebra number crunching that is not formulated in any orthogonal form even for one position leaf evaluation. Now the score is from an exhaustive search tree of leaves.
    – dbdb
    Feb 22 at 23:49

2 Answers 2

6

In short: yes. Although I don't have an overlook on current research. But already with the endgame KR/K, which is as much the drosophila of tablebases as chess is for KI, was handled this way, and for a CS project I wrote a whole miniarticle about exactly KR/K. Just the main relevant reference: look for Herbeck/Barth. (Paid access, any university access works though.) Honk if you can't get it or like more references/my article.

2
  • Github might be a friend to diffusion of knowledge for CS projects. Honk honk!
    – dbdb
    Feb 22 at 23:37
  • @dbdb: Chess Programming Wiki is up again: chessprogramming.org should have all you need. Feb 23 at 8:31
0

I the endgame table base process of production already artificial intelligence? I know of NN embeddings via LC0 architecture training in supervised mode from the EGTB results given a well controlled sampling of the input space, the sampling is covering and slicing some input space with some target covering criteria, to avoid sampling biases. However, when I was explained such experiments, I was not yet ready to fully understand the well formed questions. So what was the ambient input space in the first place (A question I am actually building right now in another thread, this question, was offered as possibly relevant. I am also curious about other answers for the op question.

I am posting here, because one might have different ways of explaining than just having a language machine working on this. I guess AI needs defining, and explanation as well. I think before EGTB gets human adapted by machines, we would need to know what makes it look "random" to a human. For now I see only humans having done that sort of work. It means research, and people asking such questions as here as starters might help gathering what it would look like. I am skeptical of just placating explainable and AI (together or not), before we ourselves can formulate what an explanation should look like.

Back to first paragraph. The most human design of chess engine I know of, that can model our thinking on the visual processing side the best, and also on its intuitive side (both together) is the A0-LC0 species. But it would have to be about data analysis of the training sets and the corresponding latent spaces of NN already trained, or during training as well.

That would be one research avenue. An automatic way to define emerging patterns. There are more constructive ways to develop explanation languages for all sort of chess. But that would also be research. I have seen some work in that direction. Even known to human recipes, like for the KNBk, could be discussed as optimal absolute or more human intuitive. Likely a definition of explainable from the mates the op is talking about would have to be able to figure out, what is more human. Rationales for the moves suggested, or equivalent less optimal solutions that might share some parts of the optimal solutions. Not all chess was tested and would make it to rule change so it would all be enjoyable, i.e. of human will to solve as individual problem to compete over in real time.

There should be some chess science interdisciplinary departments in academia. Currently I wonder if chess, as an exemplar scientific problem is not falling in the cracks of existing and possibly speciated scientific disciplines of academia or industry of AI. A question, how is one to know?

2
  • I think one straightforward approach to defining 'explanation' is to keep things in the realm of the algorithmic: an explanation is an algorithm which only needs working memory (substantially) less than the storage of the tablebase itself. For instance, it would not be difficult (is not difficult!) to program a computer to play the KBNk endgame — not perfectly, perhaps, but near enough to so as to be guaranteed to mate. There may be exceptions to the rules, but surely those exceptions will take less space than the table itself (approximately 16x55x62x61 or 3 million positions). Feb 23 at 2:16
  • I should find some lichess study links from a lichess user about just that KNBk. Even perfect might be abused to mean shortest, while as long as with N move rule, a forced mate that is longer than another is equally perfect to that other. The usual human extra rule of thumb that shorter is better might conflict with more confortable is better. That for example knight moves hopping in zig zags, might make a shorter mate, but a spatial longer bishop heavier solution, could also fence in the defending king.
    – dbdb
    Feb 28 at 3:23

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.