Talk:kd-tree

From Robowiki
Revision as of 16:21, 3 September 2009 by Zyx (talk | contribs) (→‎Tree Iterator: nice idea)
Jump to navigation Jump to search

Bucket PR k-d tree

I think you all know about Bucket PR k-d tree (aka Simonton's tree), right? I'll get directly to my point.

The Bucket PR k-d tree (I'll refer as just a tree from now on) is a binary tree split by the median of the bound, right? I wonder if I make it m-ary tree, let say, Ternary or Quaternary tree instead of just binary? Will it better? » Nat | Talk » 09:12, 15 August 2009 (UTC)

Note, it doesn't by definition have to be split by the median, though the median is usually considered the best place. As far as your second question... it may be better, or it may not. I'm pretty sure it will depend heavily on the implementation details and the data set in question. --Rednaxela 14:51, 15 August 2009 (UTC)

In a binary tree you do 1 comparison to narrow it down to 1/2 the tree (if it's well balanced). In a 3-ary you do (on average) 1.5 comparisons to narrow it to 1/3. 4-ary 2 comparisons for 1/4 of the tree. 1*1/2 = 1.5*1/3 = 2*1/4. They're all theoretically the same. --Simonton 15:33, 15 August 2009 (UTC)

  • Also note that with a 4-ary tree it's best to do a binary search at each node, which is pretty much the same thing as going back to a binary tree, except you'd be doing two comparisons for the same dimension instead of stepping to the next. --Simonton 15:36, 15 August 2009 (UTC)

(Edit conflict) It could be faster splitting 4-ways, sure, but my gut says not by very much. I'd probably recommend doing the normal 2-way (binary) one first, since it will be much simpler to develop and to debug, and then you can try to modify it to do 4-way. You'll need the 2-way version, anyway, in order to compare the speeds. Good luck. --Voidious 15:38, 15 August 2009 (UTC)

It seems that Ternary tree is the fastest of binary, ternary and quaternary. But I've some question, does the Kd-Tree output the same as linear (brute-force) search. Some of my test state that:

Starting Bucket PR k-d tree performance test...
Generating points...
Data generated.
Performing linear search...
Linear search complete; time = 0.003425156
Performing binary k-d tree search...
Binary tree search complete; time = 0.001995574
: accuracy = 0.4666666666666667
Performing ternary k-d tree search...
Ternary tree search complete; time = 2.91098E-4
: accuracy = 1.0
Performing quaternary k-d tree search...
Quaternary tree search complete; time = 3.22178E-4
: accuracy = 1.0

Data completely random. Accuracy calculate by number of result that is same as linear / cluster size. This test with a hundred data points and cluster size of 15. If I increase the data points to 26,000 then the accuracy drop to zero. Is this my tree problem or it is known problem with Kd-Tree? » Nat | Talk » 16:34, 15 August 2009 (UTC)

The kd-tree should definitely give the same results as a brute force search, so you must still have some bugs to work out. --Voidious 16:38, 15 August 2009 (UTC)

Really? Simonton's one sometimes wrong too! » Nat | Talk » 11:28, 16 August 2009 (UTC)
Well, more test with m-ary tree and Simonton's and it seems that my tree and Simonton's have the exact same output but my linear didn't. I'll try your, Rednaxela's and Chase-san's one before conclude. Expected some kd-tree benchmarks this night (ICT), afternoon (UTC) or morning (EST) » Nat | Talk » 11:56, 16 August 2009 (UTC)
Really? I'd be extremely surprised if Simonton's kd-tree had a bug in it. Maybe it's your linear search that has a bug instead (or also)? But I can say with 100% certainty that a kd-tree nearest neighbors search should produce the exact same results as a brute force nearest neighbors search. --Voidious 15:01, 16 August 2009 (UTC)

Your tree is the only tree on this site (exclude Chase-san's tree because his one doesn't have k-nearest neighbour search) that is perfect, see this:

RESULT << k-nearest neighbours search with flat/linear searching >>
: Used time             = 447.54298 x 10^{-3} seconds
: Average adding time   = 1.374 x 10^{-6} seconds
: Last node adding time = 2.4440000000000004 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Rednaxela's k-d tree >>
: Used time             = 1265.16806 x 10^{-3} seconds
: Average adding time   = 33.005 x 10^{-6} seconds
: Last node adding time = 19.696 x 10^{-6} seconds
: Accuracy              = 86%

RESULT << k-nearest neighbours search with Simonton's Bucket PR k-d tree >>
: Used time             = 96.38627 x 10^{-3} seconds
: Average adding time   = 3.039 x 10^{-6} seconds
: Last node adding time = 3.423 x 10^{-6} seconds
: Accuracy              = 80%

RESULT << k-nearest neighbours search with Nat's Bucket PR k-d tree >>
: Used time             = 117.65993 x 10^{-3} seconds
: Average adding time   = 3.368 x 10^{-6} seconds
: Last node adding time = 3.282 x 10^{-6} seconds
: Accuracy              = 80%

RESULT << k-nearest neighbours search with Voidious' Bucket PR k-d tree >>
: Used time             = 90.04413000000001 x 10^{-3} seconds
: Average adding time   = 4.639 x 10^{-6} seconds
: Last node adding time = 3.562 x 10^{-6} seconds
: Accuracy              = 100%

Actually, mine and Simonton's got all answers corrected, too, but I have this code in the answer checker:

if (Math.abs(linearAnswer[0].getDistance() - currentAnswer[0].getDistance()) > 0.00001)
	accuracy *= 0.8;

So they decrease to 80%. Input data for above result is quite large, k = 150 with 400000 data points with 15 dimensions. Here is another result with k = 5, points = 1000 and 4 dimensions:

RESULT << k-nearest neighbours search with flat/linear searching >>
: Used time             = 9.28114 x 10^{-3} seconds
: Average adding time   = 3.738 x 10^{-6} seconds
: Last node adding time = 7.613 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Rednaxela's k-d tree >>
: Used time             = 3.67735 x 10^{-3} seconds
: Average adding time   = 22.053 x 10^{-6} seconds
: Last node adding time = 65.302 x 10^{-6} seconds
: Accuracy              = 60%

RESULT << k-nearest neighbours search with Simonton's Bucket PR k-d tree >>
: Used time             = 2.47867 x 10^{-3} seconds
: Average adding time   = 8.007 x 10^{-6} seconds
: Last node adding time = 4.959 x 10^{-6} seconds
: Accuracy              = 80%

RESULT << k-nearest neighbours search with Nat's Bucket PR k-d tree >>
: Used time             = 3.32095 x 10^{-3} seconds
: Average adding time   = 5.402 x 10^{-6} seconds
: Last node adding time = 5.238 x 10^{-6} seconds
: Accuracy              = 80%

RESULT << k-nearest neighbours search with Voidious' Bucket PR k-d tree >>
: Used time             = 0.21316000000000002 x 10^{-3} seconds
: Average adding time   = 4.586 x 10^{-6} seconds
: Last node adding time = 3.5620000000000003 x 10^{-6} seconds
: Accuracy              = 100%

Note [1]: This test suite was completely rewritten. The linear search is quite slow since I just basically Collections.sort() the points.
Note [2]: My tree is 6-ary tree with 22 buckets, which is the faster combination. The fastest isn't reveal yet that is why I create this test suite » Nat | Talk » 15:35, 16 August 2009 (UTC)

Hmm, interesting. I don't understand the part where you multiply by 0.8, even though you say they got all results correct. It looks like you're just comparing the distances on the first element - is it just because those two kd-trees don't sort the results before returning them? I also don't understand how my tree could have a total time of 1/10th of the others in that last test, though I wish it were true. =) --Voidious 15:49, 16 August 2009 (UTC)

Dunno why 0.8. But, when I change my answer checking code a bit, I accidentally remove Arrays.sort() so, yeah. They are 100% now, except Rednaxela's. About the time, yes it is true. And in first test it always stay there. Mine and Simonton's usually go between 60 and 300 × 10-3. Mine usually take more time, though. Note [3]: Data completely random. » Nat | Talk » 15:56, 16 August 2009 (UTC)

Oh, and I think that's why Diamond run faster than Dookious. Separate data from the tree seems to be a good idea, especially if you set your HashMap's density to 0.1 or something. » Nat | Talk » 16:01, 16 August 2009 (UTC)

Er.... this is strange. The test code I had showed my implementation having perfect accuracy... let me test again... --Rednaxela 16:16, 16 August 2009 (UTC)

Ugh.... I found part of why your benchmark shows it so slow.... because of how you put:
public HyperPoint getPosition() {
  return new HyperPoint(location);
}

in the test code for mine... which would be a huge cause for slowness... --Rednaxela 16:33, 16 August 2009 (UTC)

After fixing that, and removing the weird 0.8 thing, I get results like:

RESULT << k-nearest neighbours search with flat/linear searching >>
: Used time             = 19.19909 x 10^{-3} seconds
: Average adding time   = 4.426 x 10^{-6} seconds
: Last node adding time = 8.032 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Rednaxela's k-d tree >>
: Used time             = 4.16715 x 10^{-3} seconds
: Average adding time   = 46.03 x 10^{-6} seconds
: Last node adding time = 9.987 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Simonton's Bucket PR k-d tree >>
: Used time             = 14.88995 x 10^{-3} seconds
: Average adding time   = 11.440000000000001 x 10^{-6} seconds
: Last node adding time = 9.289000000000001 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Nat's Bucket PR k-d tree >>
: Used time             = 5.1635800000000005 x 10^{-3} seconds
: Average adding time   = 38.910000000000004 x 10^{-6} seconds
: Last node adding time = 10.337000000000002 x 10^{-6} seconds
: Accuracy              = 100%

RESULT << k-nearest neighbours search with Voidious' Bucket PR k-d tree >>
: Used time             = 0.28069000000000005 x 10^{-3} seconds
: Average adding time   = 9.144 x 10^{-6} seconds
: Last node adding time = 8.520000000000001 x 10^{-6} seconds
: Accuracy              = 100%

most of the time, but if I re-run it, sometimes mine gets lower scores. I'll try and figure out what is happening with that. Haha, my adding speed really is terrible compared to my searching speed. --Rednaxela 16:43, 16 August 2009 (UTC)

Benchmarking adding time is really useful, isn't it? » Nat | Talk » 15:21, 17 August 2009 (UTC)

Hmm.. alright... I'm now working on a new k-d tree that should be even more efficent than the Voidious one I believe... :) --Rednaxela 23:04, 16 August 2009 (UTC)

Curses... my attempt to make a better kd-tree by using heap-style storage of the tree has failed. While it would be efficent, the kd-tree is insufficently balanced and too sparse, causing the size of array that needed to be allocated to be HUGE :( --Rednaxela 03:30, 17 August 2009 (UTC)

How much do you define HUGE? If it is ~200MB, it is acceptable, at least for me =) » Nat | Talk » 15:21, 17 August 2009 (UTC)
Oh... as in... it causes Java to run out of memory with default settings, at just 4000 data points... that's what I mean by huge :) --Rednaxela 15:25, 17 August 2009 (UTC)

(Continue on User talk:Nat/k-d tree benchmark)


Well, from some experiment with my slow m-ary k-d tree, I found that m-ary is slightly to much faster depends on number of dimension due the number of points to considered. But the slowest point of m-ary tree that make this kind of tree slower than binary k-d tree in most cases is the recursion part. Say 10-ary tree and the closest children to the center is the 6th child. Best way to do recursion is by the order 6,5,7,4,8,3,9,2,10,1. But the if/for that doing that is very expensive, it can be slower than my old, very-unoptimized linear search sometimes (6-ary with 22 bucket size, binary with 8 buckets would be up to 5 times slower). So until we can come up with better solution except hard-coded (switch(childIndex) case 1: order = new int[]{1,2,3,4,5,6,7,8,9,10} ... ), I'd say that binary is the best solution now. » Nat | Talk » 14:25, 28 August 2009 (UTC)

Hm yeah... that aspect is tricky. As far a speed though, I'm really quite doubtful that a m-ary tree can be faster than plain binary tree, at least when the binary tree is highly optimized and is smart about dimensions to split on. I'm pretty sure the asymptotic bounds on time are the same, but the well-optimized binary tree requires considerably simpler logic (like you just give) to to determine the flow of the program. I have a hunch that the only reason it's faster for you in some circumstances, is due to needing to recurse fewer times to hit the same number of nodes. I could try hacking a 3-ary version of my efficient kd-tree up, but I strongly doubt it would do any better. --Rednaxela 14:44, 28 August 2009 (UTC)

Spelling

Hmm... How this spell? kd-tree or k-d tree? Wikipedia spell it as kd-tree. But the real name is k-dimensional tree so I think it should be k-d tree. » Nat | Talk » 02:10, 23 August 2009 (UTC)

I'd say either "kd-tree" or "k-d tree" is okay. Google indicates that either is just as valid. Another note, is that if you capitalize it as in a title, it's "k-Dimensional Tree" with a lowercase k, therefore the shortened name would be either "k-D Tree" or "kD-Tree". This article should have the capitalization of 'k' and 'd' swapped :) --Rednaxela 03:35, 23 August 2009 (UTC)
Actually... I'm not sure about the capitalization of the shortenings. I see "Kd-Tree", "kD-Tree", "KD-Tree" and "KD-tree" all used. I suspect... that nobody anywhere really knows :) --Rednaxela 03:41, 23 August 2009 (UTC)
Remember that MediaWiki automatically capitalize the first letter of page. Wikipedia use {{displaytitle:}} parser hook to fix it. I think it is a time to write our own kd-tree page » Nat | Talk » 03:50, 23 August 2009 (UTC)

Tree Iterator

From what I can tell, ABC, in Shadow, does a lot of unnecessary work in pulling clusters that are bigger than necessary so that some of the results can be eliminated by out of bounds checking of the Play It Forward algorithm. How about instead having a 'tree iterator' which is initialized with a point to find neighbors to, and progressively finds the next cluster point as .next() is called? It simply stores the state of the search between calls to .next() so that the search for the next point can be resumed the moment .next() is called again. --Skilgannon 10:44, 3 September 2009 (UTC)

Hmm... I think it isn't impossible but the tree will be far more complicated. Issue that may raise is about the left/right child. Say 2 dimensions, bucket size of 1: [0.49,0] [0.51,0] [0.51,1] ... and you request 2 neighbours with center of [0.51,0]. Say it split on 0.5, on the root it will cluster first point to the left child and other two in right child. The tree will recursive into right child first and add its 2 data points into the result heap/pq/list/whatever you want. Now when it recursive into left child it will find that the data on left child is closer to center than the third data which is on the right child of the root node. If the tree go very deep this can cause a lot problem with tree iterator... Hope you understand what I'm trying to point. » Nat | Talk » 11:50, 3 September 2009 (UTC)

That's certainly an interesting thought Skilgannon. It really wouldn't be that much more complicated, however I do fear that it would be less efficent really. All it would take is storing the results in a min-heap instead of a max-heap, always calculating the distance to the path not taken, and as soon as it gets to a leaf, checking if any path not taken could possibly have something closer. That does however force extra calculations. Doing that would require an ugly amount of code-duplication as well. I might make an alternative version of the tree that supports this but I'm not sure it's worth keeping in the main version... Unless... it's somehow so efficent that I might as well just call 'next' n times to get the nearest n neighbours, but I doubt it'll be that good no matter how well I optimize it. It might be worth a try though... --Rednaxela 13:00, 3 September 2009 (UTC)

Well... after giving it some more thought... I think I may just be able to implement it without measurable efficency loss... I'll give it a try when i get back home later today :) --Rednaxela 13:24, 3 September 2009 (UTC)

That would be such a nice tree, what I like about this idea is that we wouldn't be forced to use a static cluster size, if the last processed point is close enough to the center we could keep pulling values. --zyx 15:21, 3 September 2009 (UTC)