带变点的PyMC3回归-Python问题

PyMC3 regression with change point(带变点的PyMC3回归)

本文介绍了带变点的PyMC3回归的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！

问题描述

我看到了如何使用pymc3进行变点分析的示例，但似乎遗漏了一些东西，因为我得到的结果与真实值相差甚远。这里有一个玩具示例。

数据：

脚本：

from pymc3 import *
from numpy.random import uniform, normal

bp_u = 30 #switch point
c_u = [1, -1] #intercepts before and after switch point
beta_u = [0, -0.02]  #slopes before & after switch point

x = uniform(0,90, 200)

y = (x < bp_u)*(c_u[0]+beta_u[0]*x) + (x >= bp_u)*(c_u[1]+beta_u[1]*x) + normal(0,0.1,200)

with Model() as sw_model:

    sigma = HalfCauchy('sigma', beta=10, testval=1.)

    switchpoint = Uniform('switchpoint', lower=x.min(), upper=x.max(), testval=45)

    # Priors for pre- and post-switch intercepts and slopes
    intercept_u1 = Uniform('Intercept_u1', lower=-10, upper=10)
    intercept_u2 = Uniform('Intercept_u2', lower=-10, upper=10)
    x_coeff_u1 = Normal('x_u1', 0, sd=20)
    x_coeff_u2 = Normal('x_u2', 0, sd=20)

    intercept = switch(switchpoint < x, intercept_u1, intercept_u2)
    x_coeff = switch(switchpoint < x, x_coeff_u1, x_coeff_u2)

    likelihood = Normal('y', mu=intercept + x_coeff * x, sd=sigma, observed=y)

    start = find_MAP() 

with sw_model:
    step1 = NUTS([intercept_u1, intercept_u2, x_coeff_u1, x_coeff_u2])
    step2 = NUTS([switchpoint])

    trace = sample(2000, step=[step1, step2], start=start, progressbar=True)

结果如下：

如您所见，它们与初始值有很大不同。我做错了什么？

推荐答案

最终看来，使用Metropolis采样切换到离散断点可以解决问题。以下是最终模型：

with Model() as sw_model:

    sigma = HalfCauchy('sigma', beta=10, testval=1.)

    switchpoint = DiscreteUniform('switchpoint', lower=0, upper=90, testval=45)

    # Priors for pre- and post-switch intercepts and slopes
    intercept_u1 = Uniform('Intercept_u1', lower=-10, upper=10, testval = 0)
    intercept_u2 = Uniform('Intercept_u2', lower=-10, upper=10, testval = 0)
    x_coeff_u1 = Normal('x_u1', 0, sd=20)
    x_coeff_u2 = Normal('x_u2', 0, sd=20)

    intercept = switch(switchpoint < x, intercept_u1, intercept_u2)
    x_coeff = switch(switchpoint < x, x_coeff_u1, x_coeff_u2)

    likelihood = Normal('y', mu=intercept + x_coeff * x, sd=sigma, observed=y)

    start = find_MAP() 

    step1 = NUTS([intercept_u1, intercept_u2, x_coeff_u1, x_coeff_u2])
    step2 = Metropolis([switchpoint])

    trace = sample(20000, step=[step1, step2], start=start, njobs=4,progressbar=True)

这篇关于带变点的PyMC3回归的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持编程学习网！