I need to find a model which best fits my data. It looks like this:

enter image description here

So I thought about logarithmic regression.

But when I try to make a simple fit in python I get the following result:

enter image description here

My code for now looks like this:

import csv
import numpy as np
import matplotlib.pyplot as plt
from datetime import datetime
from scipy.optimize import curve_fit
from pylab import rcParams
rcParams['figure.figsize'] = 20,10


# read the data
with open('car-2015-10-16-12-19-23.log.csv','r') as f:
    next(reader, None)
    data=np.array([tuple(row[0:]+row[:1]) for row in reader],dtype=None)

# print(mc.report_memory())   

# to test some time-windows
#data = data[500:1500]

# delete Fuel Status because sometimes there is NODATA or garbage
data = np.delete(data,np.s_[::5],1)

# convert last index to microseconds
for dt in data:
    ms = datetime.strptime(dt[-1], '%H:%M:%S.%f')
    dt[-1] = ms.microsecond + ms.second * 1000000 + ms.minute * 60 * 1000000 + ms.hour *3600 * 1000000
    dt[1] = float(dt[1]) * 1.60934

# font style
labelfont = {
        'family' : 'sans-serif',  # (cursive, fantasy, monospace, serif)
        'color'  : 'black',       # html hex or colour name
        'weight' : 'normal',      # (normal, bold, bolder, lighter)
        'size'   : 36,            # default value:12

titlefont = {
        'family' : 'serif',
        'color'  : 'black',
        'weight' : 'bold',
        'size'   : 40,

# delete garbage
data = np.delete(data, 0, 0)
data = np.delete(data, 0, 0)

# title and labels
plt.title('Throttle - Load Relation', fontdict=titlefont) 
plt.xlabel('Throttle in %', fontdict=labelfont)
plt.ylabel('Load in %', fontdict=labelfont)

# adjust fontsize of ticks
plt.tick_params(axis='both', which='major', labelsize=30)
plt.tick_params(axis='both', which='minor', labelsize=30)

# return data as float
data = data.astype(float)

# just for regression
xdata = data[:,2]
ydata = data[:,3]

# logarithmic function
def func(x, p1,p2):
  return p1*np.log(x)+p2

popt, pcov = curve_fit(func, xdata, ydata,p0=(1.0,10.2))

# curve params
p1 = popt[0]
p2 = popt[1]

# plot curve
plt.plot(curvex,curvey,'r', linewidth=5)

# plot data
plt.plot(data[:,2],data[:,3],'x',label = 'Xsaved')


The point is that both x and y can only be max 100% therefore I decided to try with logarithmic regression.

If you want you can get the data from



To better explain what I'm looking for:

enter image description here

Where btw. how to start here then?

  • 4
    $\begingroup$ What is your question? Looking at your code you may want to use pandas.read_csv instead. It's a lot easier to use than the csv module. $\endgroup$ – dsaxton Jan 10 '16 at 22:23
  • $\begingroup$ my question is if I could fit the curve more... I think it's not the best fit or my model is not enough... $\endgroup$ – x4k3p Jan 10 '16 at 22:29
  • 2
    $\begingroup$ You could try fitting a spline or loess curve. en.wikipedia.org/wiki/Scatterplot_smoothing $\endgroup$ – dsaxton Jan 10 '16 at 22:31
  • $\begingroup$ just tried with polynomial of 4th degree and got better fit. but I'm sceptical if I'm right $\endgroup$ – x4k3p Jan 10 '16 at 23:43
  • 2
    $\begingroup$ It looks to me like you might consider splines; natural regression splines might work well (with knots about every 5% at the left end where there's lots of data and maybe every 10% above x=40-ish); alternatively cubic smoothing splines might work well enough. $\endgroup$ – Glen_b Jan 11 '16 at 3:42

(Post edited based on comments - thank you for the corrections)

Your data looks very much like the data I see every day as a biochemist. I see no reason that your data should look like enzyme reaction curves, but it seems that your data may be modeled well by a fit that works well for enzyme reaction curves, generalized logistic function: https://en.wikipedia.org/wiki/Generalised_logistic_function

(Since you appear to be using python, here is a modified form of the generalized logistic, expressed in a form usable for python scripting)

def fivepl(x, a, b, c, d, g):

return ( ( (a-d) / ( (1+( (x/c)** b )) **g) ) + d )

In this variant of the generalized logistic, here are what the variables represent:

$a$ the lower asymptote

$b$ the Hill coefficient, i.e. the steepness of the slope in the linear portion of the sigmoid

$c$ is related to the value $Y(0)$, and is the inflection point of the curve, i.e. the $x$ value of the middle of the the linear portion of the curve

$d$ the upper asymptote

$g$ asymmetry factor - set to 0.5 initially

| cite | improve this answer | |
  • $\begingroup$ Looks helpful, but it's customary to state the software you are using and not to assume that people use whatever you use. I guess the last right parenthesis closing return(). comes too early. The implication presumably is to reach for some nonlinear least squares function; yours being curve_fit. $\endgroup$ – Nick Cox Apr 4 '18 at 8:33
  • $\begingroup$ I guess you're using Python like the OP. Convention here is to focus on statistics, not code in any particular software. $\endgroup$ – Nick Cox Apr 4 '18 at 8:41

Not knowing which physical process is working in the background. I would suggest to use two functions which cover different areas of the data.

enter image description here

My python skills are not sufficient to solve this task nicely, but maybe this is a beginning. I fittet your proposed function to two sections of the data. But (1) I do not know how to blend the functions nicely and (2) if it would be possible to plug this approach in the optimiser as well.

enter image description here

However, maybe another problem is the distribution of data points. Since you have a lot more data points for the low throttle area the fitting algorithm might weigh this area more (how does python fitting work?).

To prevent this I sliced the data up into 15 slices average those and than fit through 15 data points.

Averaging the slices and tweaking the model (by scaling and moving the data, be aware that most of the functions look very nice in the range from 0 to 1, but real data usually lives somewhere else.)

p1 * np.log( (x-p3)/p4 ) + p2

I end up with:

enter image description here

The green dots are the slice-averages. The parameters are:

[18.26324409 -34.75603362  15.80303842   0.10152119]
| cite | improve this answer | |
  • $\begingroup$ could anyone help me with my figure. I don't really know why they look so crappy/blurry $\endgroup$ – rul30 Mar 25 '17 at 17:30

Maybe a bit late, but this helped me:

instead of p1*np.log(x)+p2, try:

def func(x, a, b,c):
    return a*np.log2(b+x)+c

without the parameter b, I get the same problem than you but with it, it fits well.

This may then be used with scipy's curve fit:

popt, pcov = curve_fit(func, x, y)

And plotted

plt.scatter(x, y, label="Original Noised Data")
plt.plot(x, func(x, *popt), 'r-', label="Fitted Curve")

You can print popt to get the values of a,b,c.

| cite | improve this answer | |
  • 9
    $\begingroup$ Could you expand your answer a little bit more? Why do you suggest using such function? $\endgroup$ – Tim Jul 6 '16 at 12:48
  • $\begingroup$ @Sam What's your source/ref for the model design? $\endgroup$ – Rene Duchamp Feb 5 '19 at 19:19

To me that looks like a square root function, If you use a regression with the order of 1/2 that might give you a greater correlation.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.