There is more information recently that AlphaZero has been trained to be the best chess program after 4 hours of learning (in chess). I am wondering how the AI network could have been modeled for this program?

  • 3
    $\begingroup$ Could you please revise your question in line with community guidelines. $\endgroup$ – quintumnia Jan 7 '18 at 17:33
  • $\begingroup$ I think what you're really trying to ask is "What is the structure of AlphaZero?" or "What type of AI methods are involved in Alpha Zero?". Have you done the basic research yet? Looked for papers on Alpha Zero? It might be better to read as much as you can, and then revise the question to be more specific. $\endgroup$ – DukeZhou Jan 7 '18 at 21:26
  • $\begingroup$ Also worth mentioning that this person now works at DeepMind, so I might recommend his GiraffeChess paper as worthwhile. $\endgroup$ – DukeZhou Jan 7 '18 at 21:31
  • $\begingroup$ @DukeZhou No really. While GiraffeChess used neural networks, it's still a alpha-beta program. The algorithm was much worse than the Stockfish engine. The AlphaZero paper would be a better read. $\endgroup$ – HelloWorld Jan 8 '18 at 1:01
  • $\begingroup$ @SmallChess I don't disagree, but Lai wrote a statement about discontinuing Giraffe due to trade secret issues re: DeepMind's work, so there's definitely some kind of overlap. (Admittedly, this is not my area, and I just like to promote Lai's accomplishment, b/c it was steamrolled by AlphaGo a few months later.) $\endgroup$ – DukeZhou Jan 9 '18 at 0:37

This question is better on https://chess.stackexchange.com/. I've answered some of them. There's no reason to repeat here.

Please ask a specific question if you have any. I've read the paper, but I can't do anything unless you make yourself clear what you want.

| improve this answer | |

I am not familiar with AI so my idea could be basic:

  1. Input: list of booleans indicating if given field contains given checker (e.g. isA1MyRook, isA1MyKnight, isA1MyBishop, ..., isA2MyRook, isA2MyKnight...)
  2. Output: List of all possible moves including forbidden ones (e.g. Ra1-a2, Ra1-a3 etc.) I thinks that this is far from optimal but I think that this is a good starting point.
| improve this answer | |
  • 1
    $\begingroup$ I don't think this is suitable as an answer, because you preface it with "I am not familiar with AI", thus, you're just guessing. My recommendation is to do some basic research on the topic, and revise the question accordingly. If you're going to self answer, do the research first, and be confident in your conclusions! $\endgroup$ – DukeZhou Jan 7 '18 at 21:28
  • $\begingroup$ @DukeZhou. This answer is just wrong. Please revise my answer. DeepMing is much more advanced than list of booleans. $\endgroup$ – HelloWorld Jan 8 '18 at 0:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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