Permalink
Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
145 lines (98 sloc) 5.85 KB

Specifying Priors

In Bayesian statistics, data are used to update prior distributions for all parameters to calculate posterior distributions. A basic introduction to priors and Bayesian statistics can be found in "Kruschke, J. (2014). Doing Bayesian data analysis: A tutorial with R, JAGS, and Stan. Academic Press.". A more advanced treatment is given in "Gelman, A., Stern, H. S., Carlin, J. B., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian data analysis. Chapman and Hall/CRC.".

ESPEI provides a flexible interface to specify priors your want to use for a variety of parameters of different sign and magnitude through the espei.priors.PriorSpec class. This section will cover how to

  1. Define flexible, built-in priors using the YAML input and ESPEI script
  2. Use custom priors programatically

Built-in Priors

ESPEI has several built-in priors that correspond to functions in scipy.stats: uniform, normal, and triangular. There is also a special (improper) zero prior that always gives $$ ln p = 0 $$, which is the default.

Each scipy.stats prior is typically specified using several keyword argument parameters, e.g. loc and scale, which have special meaning for the different distribution functions. In order to be flexible to specifying these arguments when the CALPHAD parameters they will be used for are not known beforehand, ESPEI uses a small language to specify how the distribution hyperparameters can be set relative to the CALPHAD parameters.

Basically, the PriorSpec objects are created with the name of the distribution and the hyperparameters that are modified with one of the modifier types: absolute, relative, shift_absolute, or shift_relative. For example, the loc parameter might become loc_relative and scale might become scale_shift_relative.

Here are some examples of how the modifier parameters of value v modify the hyperparameters when given a CALPHAD parameter of value p:

  • _absolute=v always take the exact value passed in, v; loc_absolute=-20 gives a value of loc=-20.
  • _relative=v gives , v*p; scale_absolute=0.1 with p=10000 gives a value of scale=10000*0.1=1000.
  • _shift_absolute=v gives , p + v; scale_shift_absolute=-10 with p=10000 gives a value of scale=10000+(-10)=9990.
  • _shift_relative=v gives , p + v*abs(p); loc_shift_relative=-0.5 with p=-1000 gives a value of loc=-1000+abs(-1000)*0.5=-1500.

Note that the hyperparameter prefixes (loc or scale) are arbitrary and any hyperparameters used in the statistical distributions (e.g. c for the triangular distribution) could be used.

YAML

Prior can be specified in the YAML input as a list of dictionaries for different parameters. Since Python objects cannot be instantiated in the YAML files, the PriorSpec can be described a dictionary of {'name': <<NAME>>, 'loc_relative': <<VALUE>>, 'scale_relative': <<VALUE>>, ...}.

Some common examples in YAML format are as follows:

# normal prior, centered on parameter, standard deviation of 0.25*parameter
prior:
  name: normal
  loc_relative: 1.0
  scale_relative: 0.25

# uniform prior from 0 to 2*parameter (or negative(2*parameter) to 0)
prior:
  name: uniform
  loc_shift_relative: -1.0
  scale_relative: 2.0

# triangular prior, centered on parameter, from -0.5*parameter to 0.5*parameter
prior:
  name: triangular
  loc_shift_relative: -0.5
  scale_relative: 1.0

Graphically, these are shown below:

Prior example

Example priors compared to initialized parameters.

As a side note: the priors in YAML files can be passed as Python dictionaries, e.g.:

# normal prior, centered on parameter, standard deviation of 0.25*parameter
prior: {'name': 'normal', 'loc_relative': 1.0, 'scale_relative': 0.5}

Additionally, different priors can be specified using a list of prior specs that match the total degrees of freedom (VV-parameters) in the system. For example, a two parameter system could use a normal and a triangular prior simultaneously:

# two priors:
# first a normal prior, centered on parameter, standard deviation of 0.25*parameter
# second a triangular prior, centered on parameter, from -0.5*parameter to 0.5*parameter
prior: [{'name': 'normal', 'loc_relative': 1.0, 'scale_relative': 0.5}, {'name': 'triangular', 'loc_shift_relative': -0.5, 'scale_relative': 1.0}]

Custom Priors

Generally speaking, a custom prior in ESPEI is any Python object that has a logpdf method that takes a parameter and returns the natural log of the probability density function for that parameter. Any distribution you can create using the functions in scipy.stats, such as norm, is valid.

A list of these custom priors can be passed to ESPEI similar to using built-in priors, but only from the Python interface (not YAML). The number of priors must match the number of parameters, but you can also mix these with the PriorSpec objects as desired.

An example of fitting two parameters using a custom gamma distributions with minima at 10 and 100, respectively.

from scipy.stats import gamma

my_priors = [gamma(a=1, loc=10), gamma(a=1, loc=100)]

from espei.espei_script import get_run_settings, run_espei


input_dict = {
    'system': {
        'phase_models': 'phases.json',
        'datasets': 'input-data',
    },
    'mcmc': {
        'iterations': '1000',
        'input_db': 'param_gen.tdb',  # must have two parameters to fit
        'prior': my_priors,
    },
}

run_espei(get_run_settings(input_dict))