#TIL pymc.Metropolis always assume symmetric proposal distribution (default to normal), therefore does not implement Hastings correction. For a custom asymmetric proposal distribution, PyMC's out-of-the-box Metropolis algorithm will give incorrect sampling. We need to write a custom step_method.

Mathematically deriving Metropolis-Hastings algorithm from Markov-Chain's detailed balance equation makes me feel like going back to school. It's thermodynamics.

Making my repository, which is my journal in studying MCMC public. Theory, equations, and interpretations as well as code and animations/visualization can be found there.

github.com/ridhoma/mont...

The simplest illustration of Markov Chain is a random walk on a graph. Given edge Qij as probability of moving from node-i to node-j, where Qii represents probability of staying at the same node-i. Simulate the random walk, and at each time step, count how many times each node has been visited

Long time no update.. Quite busy lately in the work. But here it is. Applying Markov-Chain Monte Carlo to estimate large Traveling Salesman Problem (N=100).

It's quite fascinating that a very simple algorithm relying on random search can be used to approximate an NP-Hard optimization problem.

Studying MCMC from scratch, so far everything is basically physics ???

Bayesian Analysis generally consists of 2 essentials:
1. Assumption/priors building
2. Numerical method to get the posterior

I find this course from CMU a good starting point to learn #2 fundamental levels. I will use this as my guide
gi1242.codeberg.page/cmu-math-cs-...

First post in BlueSky. This place seems like perfect to share my learning process of MCMC methods for Bayesian Analysis.

I'll start with the very basics, the first chapter of any books: Stochastically estimating pi using the circle-in-a-square model.