Metropolis–hastings algorithm
Level: Advanced (score: 4)
In statistics and statistical physics, the Metropolis–Hastings algorithm is a Markov chain Monte Carlo (MCMC) method for obtaining a sequence of random samples from a probability distribution from which direct sampling is difficult.
This sequence can be used to approximate the distribution (e.g. to generate a histogram) or to compute an integral (e.g. an expected value).
Metropolis–Hastings and other MCMC algorithms are generally used for sampling from multi-dimensional distributions, especially when the number of dimensions is high.
Metropolis Algorithm
For the purpose of illustration, you will implement a special case of the Metropolis–Hastings algorithm where the proposal function (which candidate to choose next) is symmetric (all candidates with the same distance to the current value are equally likely).
Given a function f from which we want to draws N samples to approximate this function. f should accept a single parameter x and return the function's value for x, that is f(x).
1. Initialization: Choose an arbitrary point x_0 to be the first observation in the list of samples and choose the normal (Gaussian) distribution as a proposal function that suggests a candidate for the next sample value x_next, given the previous sample value x_t.
The Gaussian distribution is centered at the current sample x_t, so that points closer to x_t are more likely to be visited next, making the sequence of samples into a random walk.
2. For each iteration t:
- Generate a candidate x_next for the next sample by drawing a random sample from the proposal distribution that is centered at the current sample x_t.
- Calculate the acceptance ratio α=f(x_next)/f(x_t), which will be used to decide whether to accept or reject the candidate (the ratio is the probability to move to x_next).
- Accept or reject:
- Generate a uniform random number u drawn from the interval 0 to 1.
- If u ≤ α, then accept the candidate x_next as next sample.
- If u > α, then reject the candidate and stay at x_t, that is x_next=x_t
Task
Your task in this exercise is to implement the Metropolis-Hastings algorithm.
Your function will be used to approximate the mean and standard deviation values for a range of common univariate distributions like the normal distribution.
You are provided with a template for the function metropolis_hastings.
You have to implement the above mentioned algorithm in Python.
The function accepts an arbitrary probability density f (will be provided, see the tests), a start value x_0 and the number of samples n_samples to draw from f.
Step one of the algorithm is already done by passing x_0 to the function (or setting it to zero by default).
x_0 is your first sample and from there you will move either to some point in the neighborhood or stay at x_0.
For this bite, you use a normal distribution as proposal distribution with a standard deviation of one and a mean value that is equal to the current sample (x_0 at the beginner and x_t afterwards).
Refer to the tests to see how the function is called and how the statistics of f are tested.
Example
To test your function, you could try to guess the mean and standard deviation of a normal distribution like in the following example.
import numpy as np
def norm_dist(x, mean, std):
"""Gaussian normal probability distribution."""
return np.exp(-0.5 * (x - mean) ** 2 / std ** 2)
samples = metropolis_hastings(f=lambda x: norm_dist(x, mean=0, std=1), x_0=-1, n_samples=100)
print(samples.mean(), samples.std())
If your implementation is right, this should return an approximation of the true mean (0) and standard deviation (1).
Hints
numpy will come handy in this exercise as you can use it to sample both from a normal distribution and from a uniform distribution.
See numpy.random for further details.
If you want to take things up a notch, you should look into scipy.stats, the module of scipy that contains a large number of probability distributions, summary and frequency statistics, correlation functions and statistical tests. For example, scipy.stats has a norm function that can be used to draw samples from a normal distribution (norm.rvs(size=100)), to calculate the probability of an x-value (norm.pdf(x)), and to calculate the cumulative probability of an x-value (norm.cdf(x)).
As I said, scipy.stats is optional and just for the curious, you can solve this bite with numpy alone.
Happy coding!
For more background on the Metropolis–Hastings algorithm check out this page.
Further Reading Material
If you are interested in the topic, I recommend Computational Statistics in Python with a lot of code to see the concepts in action!
If you want to take things up a notch, you should look into scipy.stats, the module of scipy that contains a large number of probability distributions, summary and frequency statistics, correlation functions and statistical tests. For example, scipy.stats has a norm function that can be used to draw samples from a normal distribution (norm.rvs(size=100)), to calculate the probability of an x-value (norm.pdf(x)), and to calculate the cumulative probability of an x-value (norm.cdf(x)).
As I said, scipy.stats is optional and just for the curious, you can solve this bite with numpy alone. Be aware that scipy is not available on the pybites platform, so you have to try it locally.