/
README
99 lines (67 loc) · 3.67 KB
/
README
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
Welcome to Algorithm::LinearFit::Random!
SUMMARY
-------------------------------------------------------------------------------
This module can be used to efficiently estimate a multiple linear regression in
n-dimensional space. This is typically most useful in inconsistent,
overdetermined matrices where the "best fit" line is desired.
The method it uses to do this is to randomly adjust parameters, and check if
the resulting fit is better (has a lower RMSE) than the previous version. Over
the course of several generations, the parameters converge on a solution that
produces the minimum error.
This means that the the parameters may get "stuck" in a local minima. Optimal
selection of the random adjustment function may help prevent this problem.
Since the bulk of this module is written in C, it is pretty fast. On my own
datasets, it runs over 100x faster than the pure Perl version.
USE CASE
-------------------------------------------------------------------------------
Put another way, this module can be used to find the coefficient weights of
a dataset with multiple parameters. For example, given the data set:
Source A Source B Source C Solution
6 -4 5 4.0
1 4 6 3.0
2 8 7 6.0
1 3 8 2.5
The module would attempt to calculate the coefficient weights each source,
yielding the values [1, 0.5, 0]. Applying these to the original data,
we can verify the match.
1*6 + .5*-4 + 0*5 = 4.0
1*1 + .5* 4 + 0*6 = 3.0
1*2 + .5* 8 + 0*7 = 6.0
1*1 + .5* 3 + 0*8 = 2.5
A sample script, describing this use case, is available in example.pl
PERL INTERFACE EXAMPLE
-------------------------------------------------------------------------------
Part of the magic is that all the complexity of interacting with C is
abstracted away behind a clean idiomatic-Perl interface. For example:
#!perl -w
use Algorithm::LinearFit::Random;
my $matrix = new Algorithm::LinearFit::Random( INPUTS => 3 );
$matrix->add_row( [6, -4, 5] , 4.00 );
$matrix->add_row( [1, 3, 8] , 2.50 );
my $generations = 1_000;
my $start_weights = [0, 0, 0];
# Optimize weights for smallest RMSE
my ($rmse, $end_weights) = $matrix->tune_parameters($start_weights, $generations);
GETTING STARTED
-------------------------------------------------------------------------------
Installation requires SWIG (www.swig.org) and gcc. Building this module may be
simplified by running "perl install.pl" to build the SWIG interface and compile
the shared object.
This is clearly not ideal, and should really be replaced with a Makefile.PL
script at some point in the future. Eventually this whole thing ought to be
cleaned up, with error handling and unit tests, and then put on CPAN.
For development purposes, the library can be compiled as a standalone demo.
See src/main.c for more information.
SWIG NOTES
-------------------------------------------------------------------------------
The module is primarily written in C, for efficiency when calculating large
data sets. To interface back to the Perl, it uses the SWIG libraries. More
documentation is available from:
http://www.swig.org/Doc2.0/SWIGDocumentation.html
More details on passing arrays of floats between Perl and C is at;
http://www.swig.org/Doc2.0/Library.html#Library_carrays
TODO
-------------------------------------------------------------------------------
- Create a Makefile.PL instead of our custom build script.
- Add some unit tests, improve error handling.
- Instead of having a fixed adjustment window of 0.2, make a intelligent guess.