Python:优化求解器返回非线性回归问题的初始猜测

时间:2019-02-08 03:37:34

标签: python optimization minimize non-linear-regression

以下是ODE参数最小二乘拟合的代码。使用了Python的“最小化”和“最小二乘”功能。已经尝试了不同的方法和ODE求解器/步骤(scipy ode / odeint)。这个问题已经在MATLAB中轻松解决,但是Python不断返回初始猜测。我希望您发现编码错误,否则我会对Python优化功能感到失望。 Obj显示目标(残差平方和),而ode函数(一阶)显示参数未知的方程。数据集已附加。

import numpy as np

from scipy.integrate import ode

from scipy.optimize import least_squares

from scipy.optimize import minimize

from scipy.optimize import SR1

import matplotlib.pyplot as plt

import math


Minput=np.loadtxt('C:\\Users\\Ladan\\Documents\\Python Scripts\\Python\\moisturesmoothopt.txt') 


Minput=Minput.flatten()

time=np.linspace(0,1800,901) 

A=np.zeros(3)

XC,RC,alpha=A

#bnds=([0,0,0],[Minput[0],math.inf,math.inf])

bnds=((0,Minput[0]),(0,math.inf),(0,math.inf))

def firstorder(X,time,A):


     if X>=XC:


        dX=-RC


     if X<XC:


        dX=-RC*(X/XC)**alpha

     return dX


def obj(A):


    X0=Minput[0]

   # Xpred=odeint(firstorder,X0,time,args=(A,))

    Xpred=ode(firstorder).set_integrator('vode', method='bdf', 
    order=15).set_initial_value(Minput[0],0).set_f_params(A)

    #Xpred=ode(firstorder).set_integrator('lsoda').set_initial_value(Minput[0],0).set_f_params(A)

    EPR=Xpred

    EPR2=EPR.y.flatten()

    ERRone=np.sum(np.power((EPR2-Minput),2))


    ERR=ERRone/((901-3))    # residual sum of squares deivided by dof


    return ERR


XC=1
RC=0.005
alpha=1.5

A0=[XC,RC,alpha]


Parameters=minimize(obj,A0,method='SLSQP',bounds=bnds,options={'ftol':1e-10, 
'maxiter': 1000}) 


print('parameters',Parameters)   

Minput数组的数据在线共享:

https://1drv.ms/t/s!AoVu1vtlAOiLasJxR7rzubDr8YE

1 个答案:

答案 0 :(得分:0)

尽管我在scipy中使用了较新的ode求解器,但我倾向于偏爱良好的ol'odeint函数,该函数虽然年代久远,但仍然相当可靠,并且由于我不喜欢的原因在许多情况下提供了更好的性能不太了解。无论如何,我在很大程度上重组了您的分析代码,以同时使用scipy.optimize.least_squaresscipy.integrate.odeint。已经取得了进展,但是我确实收到有关电源中无效值的某种警告。您将需要做进一步的调查,但这应该可以使您走上正确的轨道

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
from scipy.integrate import odeint
from scipy.optimize import least_squares

Minput=np.loadtxt('Data.txt').T
time=np.linspace(0,1800,901)
bnds=([0,0,0],[Minput[0],np.inf,np.inf])


def firstorder(X,time, XC, RC, alpha):
     if X >= XC:
        dX = -RC
     else:
        dX = -RC * (X/XC)**alpha
     return dX

XC=1
RC=0.005
alpha=1.5
A0=(XC, RC, alpha) 


def residuals(x0):
    Mcalc = odeint(firstorder, y0=Minput[0], t=time, args=tuple(x0))[:,0]
    return Mcalc - Minput

Mcalc = odeint(firstorder, y0=Minput[0], t=time, args=A0)[:,0]
result = least_squares(residuals, x0=A0, bounds=bnds)
print(result)
Mfit = odeint(firstorder, y0=Minput[0], t=time, args=tuple(result.x))[:,0]

plt.plot(time, Minput, label='data')
plt.plot(time, Mcalc, label='initial values')
plt.plot(time, Mfit, label='fit')
plt.legend()

打印出:

    /---/python3.6/site-packages/ipykernel_launcher.py:20: RuntimeWarning: invalid value encountered in power

 active_mask: array([0, 0, 0])
        cost: 50.571520689865935
         fun: array([ 0.00000000e+00,  8.14148814e-03,  1.61829763e-02,  2.44244644e-02,
        ...])
        grad: array([-1.18907831,         nan, -7.75389712])
         jac: array([[ 0.        ,  0.        ,  0.        ],
       [ 0.        , -2.        ,  0.        ],
       [ 0.        , -3.99999994,  0.        ],
       ...,
       [ 0.07146036,         nan,  0.1084222 ],
       [ 0.07130456,         nan,  0.10827456],
       [ 0.07114924,         nan,  0.1081272 ]])
     message: '`xtol` termination condition is satisfied.'
        nfev: 30
        njev: 12
  optimality: nan
      status: 3
     success: True
           x: array([1.0000002 , 0.00717926, 1.50000032])

enter image description here