Skip to content

Commit

Permalink
Feature prior (#130)
Browse files Browse the repository at this point in the history
* added an initial parameter sampling functionality

* add tests for startpoint sampling

* added documentation for the additional columns intialization and objectivePriorType and Parameters
  • Loading branch information
paulstapor authored and elba.raimundez committed Oct 4, 2019
1 parent f6cceae commit 0e6c0f9
Show file tree
Hide file tree
Showing 5 changed files with 596 additions and 9 deletions.
67 changes: 58 additions & 9 deletions doc/documentation_data_format.md
Original file line number Diff line number Diff line change
Expand Up @@ -308,18 +308,39 @@ Additional columns may be added.

- `priorType`

Type of prior. Leave empty or omit column, if no priors. Normal/ Laplace etc.

[**Issue #17**](https://github.com/ICB-DCM/PEtab/issues/17)

Type of prior, which is used for sampling of initial points for
a possible optimization and for the objective function. Priors which are
only used for sampling of initial starting points or only for optimization
should be specified in the additional columns `initializationPriorType` or
`objectivePriorType`, respectivly. Possible prior types are (see Extensions):

- uniform: flat prior on linear parameters
- normal: Gaussian prior on linear parameters
- laplace: Laplace prior on linear parameters
- logNormal: exponentiated Gaussian prior on linear parameters
- logLaplace: exponentiated Laplace prior on linear parameters
- parameterScaleUniform (default): Flat prior on original parameter
scale (equivalent to "no prior")
- parameterScaleNormal: Gaussian prior on original parameter scale
- parameterScaleLaplace: Laplace prior on original parameter scale

- `priorParameters`

Parameters for prior.

[**Issue #17**](https://github.com/ICB-DCM/PEtab/issues/17)
Numeric or also parameter names? (issue #17)

Parameters for prior specified in `priorType`, separated by a semicolon.
Accordingly, there are optional columns for priors which should be used for
initial point sampling or optimization only. (i.e.,
`initializationPriorParameters` and `objectivePriorParameters`, respectively)
So far, only numeric values will be supported, no parameter names.
Parameters for the different prior types are:

- uniform: lower bound; upper bound
- normal: mean; standard deviation (**not** variance)
- laplace: location; scale
- logNormal: parameters of corresp. normal distribution (see: normal)
- logLaplace: parameters of corresp. Laplace distribution (see: laplace)
- parameterScaleUniform: lower bound; upper bound
- parameterScaleNormal: mean; standard deviation (**not** variance)
- parameterScaleLaplace: location; scale

## Parameter estimation problems combining multiple models

Expand Down Expand Up @@ -436,3 +457,31 @@ Extra columns

hierarchicalOptimization: 1 if parameter is optimized using hierarchical
optimization approach. 0 otherwise.

- `initializationPriorType` (optional)

Prior types used for sampling of initial point for optimization. Uses the
entries from `priorType` as default, but will overwrite those, if
something else is specified here. For more detailed documentation, see
`priorType`.

- `initializationPriorParameters` (optional)

Prior parameters used for sampling of initial point for optimization. Uses
the entries from `priorParameters` as default, but will overwrite those, if
something else is specified here. For more detailed documentation, see
`priorParameters`.

- `objectivePriorType` (optional)

Prior types used for the objective function during optimization. Uses the
entries from `priorType` as default, but will overwrite those, if
something else is specified here. For more detailed documentation, see
`priorType`.

- `objectivePriorParameters` (optional)

Prior parameters used for the objective function during optimization. Uses
the entries from `priorParameters` as default, but will overwrite those, if
something else is specified here. For more detailed documentation, see
`priorParameters`.
Loading

0 comments on commit 0e6c0f9

Please sign in to comment.