pin*_*gul 4 c++ random math c++11
通常我使用内置的随机函数生成值,但现在我需要创建以下形式的随机分布
f(x) = k*log(x) + m
Run Code Online (Sandbox Code Playgroud)
是否可以定义自定义随机分布函数?对于我的实际模型,我有x = [1, 1.4e7), k = -0.905787102751, m = 14.913170454
. 理想情况下,我希望它能够像当前内置发行版一样工作:
int main()
{
std::mt19937 generator;
std::uniform_real_distribution<> dist(0.0, 1.0);
my_distribution my_dist(0.0, 10.0); // Distribution using f(x)
double uni_val = dist(generator);
double log_val = my_dist(generator);
}
Run Code Online (Sandbox Code Playgroud)
这是很有可能的,但它既是一个数学问题,又是一个 C++ 问题。创建伪随机数生成器的最通用方法是逆变换采样。本质上,任何 PDF 的 CDF 均匀分布在 0 和 1 之间(如果这不明显,只需记住 CDF 的值是一个概率并思考这一点)。因此,您只需采样 0 到 1 之间的随机均匀数并应用 CDF 的逆。
在你的情况下,使用 $f(x) = k*log(x) + m$ (你没有指定界限,但我假设它们在 1 和某个正数 > 1 之间)CDF 及其逆非常混乱- 我把这个问题留给你了!C++ 中的实现如下所示
double inverseCDF(double p, double k, double m, double lowerBound, double upperBound) {
// do math, which might include numerically finds roots of equations
}
Run Code Online (Sandbox Code Playgroud)
那么生成代码将类似于
class my_distribution {
// ... constructor, private variables, etc.
template< class Generator >
double operator()( Generator& g ) {
std::uniform_real_distribution<> dist(0.0, 1.0);
double cdf = dist(g);
return inverseCDF(cdf,this->k,this->m,this->lowerBound,this->upperBound);
}
}
Run Code Online (Sandbox Code Playgroud)
我完全遵循@jwimberley 的想法,并想在这里分享我的结果。我创建了一个执行以下操作的类:
(0, 1]
std::random
代码:
// sampled_distribution.hh
#ifndef SAMPLED_DISTRIBUTION
#define SAMPLED_DISTRIBUTION
#include <algorithm>
#include <vector>
#include <random>
#include <stdexcept>
template <typename T = double, bool Interpolate = true>
class Sampled_distribution
{
public:
using CDFFunc = T (*)(T);
Sampled_distribution(CDFFunc cdfFunc, T low, T high, unsigned resolution = 200)
: mLow(low), mHigh(high), mRes(resolution), mDist(0.0, 1.0)
{
if (mLow >= mHigh) throw InvalidBounds();
mSampledCDF.resize(mRes + 1);
const T cdfLow = cdfFunc(low);
const T cdfHigh = cdfFunc(high);
T last_p = 0;
for (unsigned i = 0; i < mSampledCDF.size(); ++i) {
const T x = i/mRes*(mHigh - mLow) + mLow;
const T p = (cdfFunc(x) - cdfLow)/(cdfHigh - cdfLow); // normalising
if (! (p >= last_p)) throw CDFNotMonotonic();
mSampledCDF[i] = Sample{p, x};
last_p = p;
}
}
template <typename Generator>
T operator()(Generator& g)
{
T cdf = mDist(g);
auto s = std::upper_bound(mSampledCDF.begin(), mSampledCDF.end(), cdf);
auto bs = s - 1;
if (Interpolate && bs >= mSampledCDF.begin()) {
const T r = (cdf - bs->prob)/(s->prob - bs->prob);
return r*bs->value + (1 - r)*s->value;
}
return s->value;
}
private:
struct InvalidBounds : public std::runtime_error { InvalidBounds() : std::runtime_error("") {} };
struct CDFNotMonotonic : public std::runtime_error { CDFNotMonotonic() : std::runtime_error("") {} };
const T mLow, mHigh;
const double mRes;
struct Sample {
T prob, value;
friend bool operator<(T p, const Sample& s) { return p < s.prob; }
};
std::vector<Sample> mSampledCDF;
std::uniform_real_distribution<> mDist;
};
#endif
Run Code Online (Sandbox Code Playgroud)
以下是一些结果图。对于给定的PDF,我们需要首先通过积分来解析计算CDF。
您可以通过以下演示亲自尝试一下:
// main.cc
#include "sampled_distribution.hh"
#include <iostream>
#include <fstream>
int main()
{
auto logFunc = [](double x) {
const double k = -1.0;
const double m = 10;
return x*(k*std::log(x) + m - k); // PDF(x) = k*log(x) + m
};
auto sinFunc = [](double x) { return x + std::cos(x); }; // PDF(x) = 1 - sin(x)
std::mt19937 gen;
//Sampled_distribution<> dist(logFunc, 1.0, 1e4);
Sampled_distribution<> dist(sinFunc, 0.0, 6.28);
std::ofstream file("d.txt");
for (int i = 0; i < 100000; i++) file << dist(gen) << std::endl;
}
Run Code Online (Sandbox Code Playgroud)
数据是用python绘制的。
// dist_plot.py
import numpy as np
import matplotlib.pyplot as plt
d = np.loadtxt("d.txt")
fig, ax = plt.subplots()
bins = np.arange(d.min(), d.max(), (d.max() - d.min())/50)
ax.hist(d, edgecolor='white', bins=bins)
plt.show()
Run Code Online (Sandbox Code Playgroud)
运行演示:
clang++ -std=c++11 -stdlib=libc++ main.cc -o main; ./main; python dist_plot.py
Run Code Online (Sandbox Code Playgroud)