ControlVariates

Still under construction, but should be able to use. Any comments are welcome.

"Control variates" is a postprocess technique to reduce the variance of MCMC samples. It can achieve better approximation of expectation of random variables with small number of samples, see the example in example_cv.py.

The processing chain is:

1. Draw samples using any tool. (e.g. pystan3)

2. Generate a matrix including the samples. (e.g. pystan3samples_to_matrix() in postprocess_bs.py)

3. Extract gradient of log-probability. (e.g. using bridgestan to expose (see run_postprocess() in postprocess_bs.py))

4. Construct control variates and generate new samples. (see controlvariates_basics.py)

Requirement

• The example uses pystan3 to generate MCMC samples:

$ pip install pystan

• Any dependants that are required by pystan3.

Install

$ git clone https://github.com/zhouyifan233/ControlVariates --recursive

Bridgestan 1.0.0 will be automatically downloaded.

Run example

Configure cmdstan path in example_cv.py (line 9).

$ cd ControlVariates
$ python example_cv.py

You can try different models: just change exp_path (line 12).

Paper

Control Variates for Constrained Variables

Please consider cite our work if you find it useful:

@ARTICLE{9944852,

author={Maskell, Simon and Zhou, Yifan and Mira, Antonietta},

journal={IEEE Signal Processing Letters},

title={Control Variates for Constrained Variables},

year={2022},

volume={29},

number={},

pages={2333-2337},

doi={10.1109/LSP.2022.3221347}

}