The Simplex algorithm of Nelder & Mead is a more robust but inefficient (slow) optimisation algorithm.
It only uses function evaluations but no gradients or inferred gradients. The score function is minimised geometrically be stepping in different directions, trying different stepsizes. The Simplex is a greedy algorithm, too.
The objective function to be minimised is directly provided and does not need to be of a least-squares form. Hence, Simplex is much more general than leastsq or curve_fit.
Here is a code example:
from scipy.optimize import fmin as simplex
# "fmin" is not a sensible name for an optimisation package.
# Rename fmin to "simplex"
# Define the objective function to be minimised by Simplex.
# params ... array holding the values of the fit parameters.
# X ... array holding x-positions of observed data.
# Y ... array holding y-values of observed data.
# Err ... array holding errors of observed data.
def func(params, X, Y, Err):
# extract current values of fit parameters from input array
a = params[0]
b = params[1]
c = params[2]
# compute chi-square
chi2 = 0.0
for n in range(len(X)):
x = X[n]
# The function y(x)=a+b*x+c*x^2 is a polynomial
# in this example.
y = a + b*x + c*x*x
chi2 = chi2 + (Y[n] - y)*(Y[n] - y)/(Err[n]*Err[n])
return chi2
xdata = [0.0,1.0,2.0,3.0,4.0,5.0]
ydata = [0.1,0.9,2.2,2.8,3.9,5.1]
sigma = [1.0,1.0,1.0,1.0,1.0,1.0]
#Initial guess.
x0 = [0.0, 0.0, 0.0]
# Apply downhill Simplex algorithm.
print simplex(func, x0, args=(xdata, ydata, sigma), full_output=0)
The result (a=0.10001189, b=0.88144704, c=0.021426) has no error/uncertainty estimates because the Simplex algorithm is a mere optimisation. Given the Simplex estimate of the minimum, errors have to be estimated afterwards using some other method.