[mlpack] Numeric sensitivity of RangeSearch

[Li Dong]

Dear Ryan,

I have isolated a bug case into a single code in the attachment, from which you can see the results:

Neighbor grid number is 201 when R is 1.
Neighbor grid number is 102 when R is 6.371e+06.

Best,

Li

On Mar 25, 2014, at 8:40 PM, Ryan Curtin <gth671b at mail.gatech.edu> wrote:

> On Tue, Mar 25, 2014 at 04:35:18PM +0800, Li Dong wrote:
>> Dear all,
>> 
>> I have used RangeSearch to search neighbours on the sphere surface.
>> Previously, the radius of sphere is set to 1, so the Cartesian
>> coordinate range (I convert spherical coordinate to Cartesian
>> coordinate for searching) is from -1 to 1. In real applications, the
>> radius of the Earth is around 6.371e6, and the Cartesian coordinate
>> range is also from -6.371e6 to 6.371e6. The problem that I am facing
>> is when radius is 6.371e6, many neighbours are not found by
>> RangeSearch.
>> 
>> I have uploaded the result comparison in
>> https://gist.githubusercontent.com/dongli/9634592/raw/cb3e6cdf85758594be15fabd7fa61d73ac4ecf3e/rangesearch_numeric_problem.png,
>> where the blue circle represents the range (approximately), and red
>> points are the neighbours. The only difference is the radius.
> 
> Hi Li,
> 
> I have tried to generate a similar example, running range search first
> on a similar-looking dataset normalized to [-1, 1], and then
> renormalized to [-6e6, 6e6], but the results were the same.  Can you
> give more information on your dataset or some way to reproduce the error
> so I can look into it further?
> 
> Thanks,
> 
> Ryan
> 
> -- 
> Ryan Curtin    | "I like this game."
> ryan at ratml.org |   - Coach

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20140325/29679519/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: debug.cpp
Type: application/octet-stream
Size: 2293 bytes
Desc: not available
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20140325/29679519/attachment.obj>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20140325/29679519/attachment-0001.html>