Easy Metropolis-Hastings’s documentation!¶
Metropolis-Hastings is one of the two well-known sampling algorithms (with the other one being Gibbs) for Markov Chain Monte-Carlo (MCMC) estimation methods, which are frequently used in Bayesian statistics, Graphical models, etc.. This lightweight Python package helps you use MCMC based on Metropolis-Hastings sampling to solve your own problems in research or engineering.
Why another package?¶
There are already several powerful Python packages in this direction, one may raise the question why I bother to build another one. While those full-fledged packages are powerful, I see their strength as their own weakness. For instance, when you are in the climax of solving your research problem and you just need a single function (just like the minimize in Scipy) to score the goal. At this exciting moment where you are only one step away from your Goddess, you find yourself having to calm down and consult a one hundred page manual to find the correct way to use a package. Oh, shit…
With this package, everything becomes easy. You no longer need to climb the steep learning curve like reading a medical notice or a law contract. With merely one line of code, the famous Metropolis-Hastings algorithm is within your hand. While your peers are sweating about the error messages of a complex package, you have already solved your problem and moved on. All these advantages are brought to you by this package.
Features¶
- One-liner: It requires only one line of code to plug a MCMC method in your program.
- Purity: Everything is written in Python and open-source. There is no hidden secret.
- In-cube sampling: The sampling domain does not need to be the whole space. It can also be an open cube, bounded or not. For closed cubes, just add epsilons to the cube.
- Flexibility: Within each iteration, you can sample dimension by dimension or sample once for all dimensions. Or you can sample a single dimension during each iteration in a rotative fashion. The inner sampling distribution can be either Gaussian or uniform.
- State-variant: Your sampling strategy can depend not only on the time but also on the state.
- God Space: The current space may not be an ideal place to apply MCMC, and some “God Space” may be a better choice. By applying a parameter transformation, you ascend to the God Space, and descend to the current space later.
- Track slicer: You can choose the indices on which to compute your final estimate. It not only gives you the way to specify the burn-in period but also allows you to pick a sub series.
- Reducer: You do not have to use the mean function to obtain the final estimate. You can also use median or any other reducer, built-in or not.
- Customizability: You can supply your own move function and/or law function to customize the behavior of the algorithm, in essentially the same way as when you use the map-reduce frameworks.