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
|
# Cross-shaped anti-correlation
# =============================
#
# Example model with strong correlations between the fitted parameters.
#
# In this case we define the probability density function (PDF) directly
# as an 'X' pattern, with width sigma.
#
# Ideally, the a-b correlation plot will show the 'X' completely filled
# within the bounds.
from bumps.names import *
# Adjust scale from 1e-150 to 1e+150 and you will see that DREAM is equally
# adept at filling the cross. However, if sigma gets too small relative to
# scale the fit will get stuck on one of the arms, and if sigma gets too
# large, then the whole space will be filled and the x will not form.
scale = 10
sigma = 0.1 * scale
# sigma = 0.001*scale # Too small
# sigma = 10*scale # Too large
# Simple gaussian cost function based on the distance to the closest ridge
# *x=y* or *x=-y*.
def fn(a, b):
return 0.5 * min(abs(a + b), abs(a - b)) ** 2 / sigma**2 + 1
# Wrap it in a PDF object which turns an arbitrary probability density into
# a fitting function. Give it an initial value away from the cross.
M = PDF(fn, a=3 * scale, b=1.2 * scale)
# Set the range of values to include the cross. You can skip the center of
# the cross by setting b.range to (1,3), and for reasonable values of sigma
# both arms will still be covered. Extend the range too far (e.g.,
# a.range(-3000,3000), b.range(-1000,3000)), and like a value of sigma
# that is too small, only one arm of the cross will be filled.
M.a.range(-3 * scale, 3 * scale)
M.b.range(-1 * scale, 3 * scale)
# Make the PDF a fit problem that bumps can process.
problem = FitProblem(M)
|
