cbayes

Note: Currently in heavy development.

A collection of Python scripts to make Bayesian inference of unifilar hidden Markov models (uHMMs) and epsilon-machines with Computational Mechanics Python, aka CMPy easier. The goal is to have an easy-to-use pipeline of scripts that maximize the use of all compute power on a single machine-- this is done using multiprocessing to launch many processes whenever possible.

Dependencies

• CMPy

CMPy should be installed with the --cython switch to allow for use of the enumeration library. Also, note that CMPy has a quite a few dependencies that are required to use these scripts.

Installation

Get the code from github in the usual manner. In a suitable directory clone the repository using:

git clone [email protected]:cstrelioff/cbayes.git

Install using:

sudo python setup.py install

Or, for a local install, when administrator permissions are not available:

python setup.py install --user

If you have issues with the scripts when using the --user switch, consult the local installation note. The scripts are designed to be run as system tools, allowing for execution from any directory without directly calling python.

Scripts

All of these scripts should install as system scripts, making use from the command line easy. To get help for a script, type

scriptname --help

High Level Scripts

The scripts produce a master script that calls each of the base scripts (see below) repeatedly to accomplish some larger inference goal. Broadly speaking there are two types of analyses:

• Convergence -- a single data series is repeatedly analyzed, starting with a small segment, and increasing the segment length until the full data set is used. This analysis is helpful for looking at convergence of inference with increasing data size.

• Overlap -- a single data series is repeatedly analyzed, using reasonable sized overlapping chunks, to test for stationary behavior. The same model topologies should probable for each segment, reflecting in-class and stationary behavior.

The bash scripts are designed to run on any Linux/Mac machine and the slurm scripts can be used on a cluster running the slurm resource manager.

cbayes_bash_enumerate_convergence.py

Create a bash script to run a sequence of cbayes_enumerate_ scripts on a specified data file. In this case, the focus is on convergence of inference by considering segments, using the single data file, of increasing length.

cbayes_bash_enumerate_overlap_analysis.py

Creates a bash script to run a sequence of cbayes_enumerate_ scripts on a specified data file. In this case, the focus is on analysis of single data file as well as considering overlapping subsegments to test for stationary behavior. In this context, stationary means that analysis of segments of the total data series returns the same model(s) at different points -- reflecting a single model topology, or set of topologies, is appropriate for describing the data series.

cbayes_slurm_enumerate_convergence.py

Create a slurm script to run a sequence of cbayes_enumerate_ scripts on a specified data file. In this case, the focus is on convergence of inference by considering subsegments, using the single data file, of increasing length.

cbayes_slurm_enumerate_overlap_analysis.py

Create a slurm script to run a sequence of cbayes_enumerate_ scripts on a specified data file. In this case, the focus is on analysis of single data file as well as considering overlapping segments to test for stationary behavior. In this context, stationary means that analysis of segments of the total data series return the same model at different points -- reflecting a single, static model topology is appropriate for the complete data series.

Base Scripts

These base scripts accomplish very specific tasks. While they can be used alone, it is often useful to employ the high level scripts (see above) to generate a bash or slurm master script for generation of a whole research protocol. These master scripts use base scripts described below.

cbayes_create_candidate_models.py

This script is used to create a file with all candidate machine topologies that the user wants to consider. In addition, a database directory -- really just a normal directory -- will be created for later use by other scripts.

cbayes_create_process_datafile.py

Create a data file that works nicely with other cbayes scripts. Can use any machine in cmpy.machines that has default parameters. This means the process can be instantiated using:

em = cmpy.machines.ProcessName()

Note: If you have your own data, the output of this script would be a good reference to see how the data file is formatted.

cbayes_enumerate_prior.py

This script focuses on the prior over models, calculating the prior log evidence for all model topologies in the machines file.

cbayes_enumerate_posterior.py

This script focuses on the posterior over models, calculating the posterior log evidence for all model topologies in the machines file using the specified data file and data range.

cbayes_enumerate_probabilities.py

This script processes the output of cbayes_enumerate_prior.py or cbayes_enumerate_posterior.py scripts to calculate prior, or posterior, probabilities for each machine/model topology using a specified value of beta. This script must be run before samples can be generated from the prior or posterior.

cbayes_enumerate_sample.py

This script generates samples from the prior, or posterior, over model topologies using the output a specified model probabilities file.

GA Scripts