cs0x7f / sq12phase Goto Github PK

Hey! I'm currently maintaining TNoodle and sq1 is causing us a headache, because it slows down our integration tests by a ton. The reason is that we still use the sub-optimal solver before your fix.

I understand that I can just cherry-pick your recent commits, but we still cannot use your optimal solver because of conflicts with the turn metrics:

• Your solver counts each U and D face turn individually, so (1, 2) is actually 2 moves
• WCA Regulation 12c4 requires that each (U, D) group counts as 1 move

I have considered a "hacky" solution where your program counts each U and D as 0.5 moves, but that doesn't work because (1, 0) / (2, 3) becomes 2.5 moves when it should be 3 per WCA.

Apart from this initial idea, I have failed to implement this metric myself because the code is too advanced for me 🙈
Is there any chance you could implement 12c4 metrics either here directly or as a PR on our separate sq12phase clone over at https://github.com/thewca/tnoodle-lib ?

Thanks a lot for your help! :)

I added timing of the initialization to Search.java (see https://github.com/jfly/sq12phase/commit/12752059ae7d6f3e46dc21b50b6eb1764b458bd3). I'm seeing times like

2.87255555 seconds to initialize
2.644265412 seconds to initialize
2.640455625 seconds to initialize

which is a lot slower than the nearly instantaneous speed of the js on http://www.cubing.net/mark2/#competition_name=&rounds=%5B%5B%22sq1%22%2C%22Round%201%22%2C1%2C5%5D%5D.

When toying around with extremely simple puzzle states in tnoodle-lib, the optimal solver crashes with an ArrayIndexOutOfBoundsException right at https://github.com/thewca/tnoodle-lib/blob/master/sq12phase/src/main/java/cs/sq12phase/FullCube.java#L157

An example to reproduce this is (2, -4). I have used the following debug method to figure out the puzzle state:

int[] debugState() {
    int[] debug = new int[24];

    for (int i = 0; i < 24; i++) {
        debug[i] = pieceAt(i);
    }

    return debug;
}

For the above example, TNoodle yields the following "converted" FullCube output:
[1,2,3,3,4,5,5,6,7,7,0,1,12,15,15,14,9,9,8,11,11,10,13,13]

Upon further inspection, the initial error makes sense because the piece 1 (a corner that is 2 vertices wide) "wraps around" and thus makes cnt think it's a new piece. However, I am astonished that for > 100 random "long" scrambles, this problem does not occur and I can only forcibly reproduce it for extremely short scramble sequences.

I thus have two major suspicions:
(a) The getParity code acutally has a flaw
(b) The converter from TNoodle --> sq12phase is faulty, in which case I need input from either you or @jfly concerning https://github.com/thewca/tnoodle-lib/blob/master/scrambles/src/main/java/net/gnehzr/tnoodle/puzzle/SquareOnePuzzle.java#L244

Please advise how to proceed