Thanks in advance

[sfp932705]

Hey guys!
I have used you indexing function indices(). Works great!
One question though, is there any way to use it for a 2 dimensional array?

[Korijn]

Can you share a minimal piece of code that demonstrates what you want to accomplish?

[sfp932705]

Sure.
I am implementing A* to solve N-Puzzle, and need to calculate manhattan distance between 2 states. A state is an nxn 2D array. I have done this using loops, as follows

...
for i in xrange(np.size(Mxy[0])):
Gxy[0][i] = np.where(goal == matrix[Mxy][i])[0]
Gxy[1][i] = np.where(goal == matrix[Mxy][i])[1]
Cxy = abs(Mxy[0]-Gxy[0]) + abs(Mxy[1]-Gxy[1])
return np.sum(Cxy)

where Mxy holds the tiles that are differently placed in both states, and Gxy is just an empty array with same dimensions. It works okay, but as n grows it takes more time. And i have defined a linear conflict function, that needs to calculate manhattan distance first. I want to speed things up a little bit.
I used numpy_indexed.indices(goal,matrix) for a flat array version of my states and it works great. However i need to use it in the actual two dimensional array in order to get the actual manhattan distance.
I just want to get rid of the python loop, and have found several ways but only with a flattened array. I tried the scipy cdist but didn't get what i expected.
Any help is appreciated!!!

[Korijn]

Have you tried browsing the unit tests of the indices function? https://github.com/EelcoHoogendoorn/Numpy_arraysetops_EP/blob/master/tests/test_indexed.py#L80

[sfp932705]

Actually no, but what should i do? Look through to see if any function is of interest to my approach?

[Korijn]

Yeah. I'm afraid I'm not too sure how to deal with your example code. Hopefully @EelcoHoogendoorn can answer your question directly.

[sfp932705]

ok thanks!

[Korijn]

Actually, do you mind providing an example input & output, so I can test my solution? I might be able to figure it out.

[sfp932705]

This is kind of what i want:
Given a goal state,
[[1,2,3,4]
[5,6,7,8]
[9,10,11,12]
[13,14,15,0]]
And a current state,
[[4,1,5,6]
[2,3,8,7]
[9,10,11,12]
[14,15,13,0]]

I would like to know the index for every element of the current state in the goal state, in some sort of vectorized way that involves no looping. So i would expect getting:
[[[0,3],[0,0],[1,0],[1,1]],..] and so on, because 4 which is the first element in current state is element [0,3] in goal state,and 1 is element [0,0,] and so on.
Thank you!

[EelcoHoogendoorn]

Hi;

There are various graph-like algorithms that you can reasonably efficiently implement using the techniques in this library; at least compared to any python for loops; but its not going to be very fast, compared to the very optimized graph libraries available in python for A* specifically. But certainly interesting as an exercise in vectorized graph algorithms.

It is not clear to me what the code you posted thus far is trying to accomplish though; as per stackoverflow, a self-contained example would help a lot. Speaking of which; its a better format for these type of questions.

[EelcoHoogendoorn]

Ah posted that last comment without refreshing my page; thats a lot more of a self-contained problem!

Indeed you could do something like:

idx = npi.indices(G.flatten(), C.flatten())
idx = np.unravel_index(idx, C.shape)

on a higher level, I really doubt this is the most efficient way to approach A* using numpy_indexed though...

[sfp932705]

ok thank you very much!