Estimation

ActivitySim includes the ability to re-estimate submodels using choice model estimation tools such as larch. To do so, ActivitySim adopts the concept of an estimation data bundle (EDB), which is a collection of the necessary data to re-estimate a submodel. For example, for the auto ownership submodel, the EDB consists of the following files:

  • model settings - the auto_ownership_model_settings.yaml file

  • coefficients - the auto_ownership_coefficients.csv file with each coefficient name, value, and constrain set to True or False if the coefficient is estimatable

  • utilities specification - the auto_ownership_SPEC.csv utility expressions file

  • chooser and alternatives data - the auto_ownership_values_combined.csv file with all chooser and alternatives data such as household information, land use information, and the utility data components for each alternative

ActivitySim also includes Jupyter Estimation Notebooks for estimating submodels with larch, as well as an activitysim.estimation.larch submodule that transforms EDBs into larch models. Additional estimation software translators can be added later if desired.

The combination of writing an EDB for a submodel + a larch estimation notebook means users can easily re-estimate submodels. This combination of functionality means:

  • There is no duplication of model specifications. ActivitySim owns the specification and larch pivots off of it. Users code model specifications and utility expressions in ActivitySim so as to facilitate ease of use and eliminate inconsistencies and errors between the code used to estimate the models and the code used to apply the models.

  • The EDB includes all the data and model structure information and the activitysim.estimation.larch submodule used by the example notebooks transforms the EDB to larch’s data model for estimation.

  • Users are able to add zones, alternatives, new chooser data, new taz data, new modes, new coefficients, revise utilities, and revise nesting structures in ActivitySim and larch responds accordingly.

  • Eventually it may be desirable for ActivitySim to automatically write larch estimators (or other types of estimators), but for now the integration is loosely coupled rather than tightly coupled in order to provide flexibility.

Workflow

The general workflow for estimating models is shown in the following figures and explained in more detail below.

_images/estimation_tools.jpg

  • The user converts their household travel survey into ActivitySim format households, persons, tours, joint tour participants, and trip tables. The households and persons tables must have the same fields as the synthetic population input tables since the surveyed households and persons will be run through the same set of submodels as the simulated households and persons.

  • The ActivitySim estimation example scripts\infer.py module reads the ActivitySim format household travel survey files and checks for inconsistencies in the input tables versus the model design + calculates additional fields such as the household joint tour frequency based on the trips and joint tour participants table. Survey households and persons observed choices much match the model design (i.e. a person cannot have more work tours than the model allows).

  • ActivitySim is then run in estimation mode to read the ActivitySim format household travel survey files, run the ActivitySim submodels to write estimation data bundles (EDB) that contains the model utility specifications, coefficients, chooser data, and alternatives data for each submodel. Estimation mode runs single-processed and without destination sampling.

  • The relevant EDBs are read and transformed into the format required by the model estimation tool (i.e. larch) and then the coefficients are re-estimated. The activitysim.estimation.larch library is included for integration with larch and there is a Jupyter Notebook estimation example for each core submodel. No changes to the model specification are made in the process.

  • The user can then update the ActivitySim model coefficients file(s) for the estimated submodel and re-run the model in simulation mode. The user may want to use the restartable pipeline feature of ActivitySim to just run the submodel of interest.

_images/estimation_example.jpg

Example

Note

The estimation_mode.ipynb Jupyter Estimation Notebooks also introduces estimation mode and walks the user through the process.

To run the estimation example, do the following:

  • Activate the correct conda environment if needed

  • Create a local copy of the estimation example folder

activitysim create -e example_estimation_sf -d test_example_estimation_sf

  • Run the example

cd test_example_estimation_sf
activitysim run -c configs_estimation/configs -c configs -o output -d data_sf

  • ActivitySim should log some information and write outputs to the output folder, including EDBs for each submodel. The estimation example runs for about 15 minutes and writes EDBs for 2000 households.

  • Open Estimation Notebooks for a specific submodel and then step through the notebook to re-estimate the sub-model.

The estimation example assumes the machine has sufficient RAM to run with chunking disabled (chunk_training_mode: disabled). See Chunk for more information.

Settings

Additional settings for running ActivitySim in estimation mode are specified in the estimation.yaml file. The settings are:

  • enable - enable estimation, either True or False

  • bundles - the list of submodels for which to write EDBs

  • survey_tables - the list of input ActivitySim format survey tables with observed choices to override model simulation choices in order to write EDBs. These tables are the output of the scripts\infer.py script that pre-processes the ActivitySim format household travel survey files for the example data and submodels

Estimation Notebooks

ActivitySim includes a Jupyter Notebook recipe book with interactive re-estimation examples for each estimatable submodel. To run a Jupyter notebook, do the following:

  • Open a conda prompt and activate the conda environment with ActivitySim installed

  • If needed, conda install jupyterlab so you can run jupyter notebooks

  • Type jupyter notebook to launch the web-based notebook manager

  • Navigate to the examples/examples_estimaton/notebooks folder and select a notebook from the table below

  • Save the updated coefficient file(s) to the configs folder and run the model in simulation mode

Example

Notebook

Estimation mode overview

01_estimation_mode.ipynb

School location

02_school_location.ipynb

Work location

03_work_location.ipynb

Auto ownership

04_auto_ownership.ipynb

Free parking

05_free_parking.ipynb

CDAP

06_cdap.ipynb

Mandatory tour frequency

07_mand_tour_freq.ipynb

Work tour scheduling

08_work_tour_scheduling.ipynb

School tour scheduling

09_school_tour_scheduling.ipynb

Joint tour frequency

10_joint_tour_freq.ipynb

Jointatory tour composition

11_joint_tour_composition.ipynb

Jointatory tour participation

12_joint_tour_participation.ipynb

Joint nonmandatory tour destination

13_joint_nonmand_tour_dest.ipynb

Joint tour scheduling

14_joint_tour_scheduling.ipynb

Non mandatory tour frequency

15_non_mand_tour_freq.ipynb

Non mandatory tour scheduling

16_nonmand_tour_scheduling.ipynb

Tour mode choice

17_tour_mode_choice.ipynb

Atwork subtour frequency

18_atwork_subtour_freq.ipynb

Atwork subtour destination

19_atwork_subtour_dest.ipynb

Atwork subtour scheduling

20_atwork_subtour_scheduling.ipynb

Stop frequency

21_stop_frequency.ipynb

Trip destination

22_trip_dest.ipynb

Trip mode choice

23_trip_mode_choice.ipynb

General API

activitysim.estimation.larch.general.apply_coef_template(linear_utility, template_col, condition=None)

Apply a coefficient template over a linear utility function.

Parameters
linear_utilityLinearFunction_C
template_colMapping
conditionany
Returns
LinearFunction_C
activitysim.estimation.larch.general.apply_coefficients(coefficients, model, minimum=None, maximum=None)

Read the coefficients CSV file to a DataFrame and set model parameters.

Parameters
coefficientspandas.DataFrame

The coefficients table in the ActivitySim data bundle for this model.

modelModel

Apply coefficient values and constraints to this model.

activitysim.estimation.larch.general.clean_values(values, choice_col='override_choice', alt_names_to_codes=None, choice_code='override_choice_code')
Parameters
valuespd.DataFrame
alt_namesCollection
override_choicestr

The columns of values containing the observed choices.

alt_names_to_codesMapping, optional

If the override_choice column contains alternative names, use this mapping to convert the names into alternative code numbers.

choice_codestr, default ‘override_choice_code’

The name of the observed choice code number column that will be added to values.

Returns
pd.DataFrame
activitysim.estimation.larch.general.construct_nesting_tree(alternatives, nesting_settings)

Construct a NestingTree from ActivitySim settings.

Parameters
alternativesMapping or Sequence

If given as a Mapping (dict), the keys are the alternative names as strings, and the values are alternative code numbers to use in larch. If given as a Sequence, the values are the alternative names, and unique sequential codes will be created starting from 1.

nesting_settingsMapping

The ‘NESTS’ section of the ActivitySim config file.

Returns
NestingTree
activitysim.estimation.larch.general.cv_to_ca(alt_values, dtype='float64', required_labels=None)

Convert a choosers-variables DataFrame to an idca DataFrame.

Parameters
alt_valuespandas.DataFrame

This DataFrame should be in choosers-variables format, with one row per chooser and variable, and one column per alternative. The id’s for the choosers and variables must be in that order, in a two-level MultiIndex.

dtypedtype

Convert the incoming data to this type. Set to None to skip data conversion.

required_labelsCollection, optional

If given, any columns in the output that are not required will be pre-emptively dropped.

Returns
pandas.DataFrame

The resulting DataFrame is transformed into Larch’s idca format, with one row per chooser (case) and alternative, and one column per variable.

activitysim.estimation.larch.general.dict_of_linear_utility_from_spec(spec, x_col, p_col, ignore_x=())

Create a linear function from a spec DataFrame.

Parameters
specpandas.DataFrame

A spec for an ActivitySim model.

x_col: str

The name of the columns in spec representing the data.

p_col: dict

The name of the columns in spec representing the parameters. The keys give the names of the columns to use, and the values will become the keys of the output dictionary.

ignore_xCollection, optional

Labels in the spec file to ignore. Typically this includes variables that are pre-processed by ActivitySim and therefore don’t need to be made available in Larch.

segment_idstr, optional

The CHOOSER_SEGMENT_COLUMN_NAME identified for ActivitySim. This value is ignored if p_col is a string, and required if p_col is a dict.

Returns
dict
activitysim.estimation.larch.general.explicit_value_parameters(model)

Define and lock parameters given as fixed values.

Parameters
modellarch.Model

The model to insert fixed value parameters.

activitysim.estimation.larch.general.explicit_value_parameters_from_spec(spec, p_col, model)

Define and lock parameters given as fixed values in the spec.

Parameters
specpandas.DataFrame

A spec for an ActivitySim model.

p_colstr or dict

The name of the columns in spec representing the parameters. Give as a string for a single column, or as a dict to have segments on multiple columns. If given as a dict, the keys give the names of the columns to use, and the values give the identifiers that will need to match the loaded segment_id value. Only the keys are used in this function.

modellarch.Model

The model to insert fixed value parameters.

activitysim.estimation.larch.general.linear_utility_from_spec(spec, x_col, p_col, ignore_x=(), segment_id=None)

Create a linear function from a spec DataFrame.

Parameters
specpandas.DataFrame

A spec for an ActivitySim model.

x_col: str

The name of the columns in spec representing the data.

p_col: str or dict

The name of the columns in spec representing the parameters. Give as a string for a single column, or as a dict to have segments on multiple columns. If given as a dict, the keys give the names of the columns to use, and the values give the identifiers that will need to match the loaded segment_id value.

ignore_xCollection, optional

Labels in the spec file to ignore. Typically this includes variables that are pre-processed by ActivitySim and therefore don’t need to be made available in Larch.

segment_idstr, optional

The CHOOSER_SEGMENT_COLUMN_NAME identified for ActivitySim. This value is ignored if p_col is a string, and required if p_col is a dict.

Returns
LinearFunction_C
activitysim.estimation.larch.general.prevent_overlapping_column_names(x_ca, x_co)

Rename columns in idca data to prevent overlapping names.

Parameters
x_ca, x_copandas.DataFrame

The idca and idco data, respectively

Returns
x_ca, x_co
activitysim.estimation.larch.general.remove_apostrophes(df, from_columns=None)

Remove apostrophes from columns names and from data in given columns.

Also strips leading and trailing whitespace.

This function operates in-place on DataFrames.

Parameters
dfpandas.DataFrame
from_columnsCollection, optional
Returns
pandas.DataFrame
activitysim.estimation.larch.simple_simulate.construct_availability(model, chooser_data, alt_codes_to_names)

Construct an availability dataframe based on -999 parameters.

Parameters
modellarch.Model
chooser_datapandas.DataFrame
alt_codes_to_namesMapping[int,str]
Returns
pandas.DataFrame

Models API

activitysim.estimation.larch.cdap.apply_replacements(expression, prefix, tokens)

Convert general person terms to specific person terms for the CDAP model.

Parameters
expressionstr

An expression from the “Expression” column of cdap_INDIV_AND_HHSIZE1_SPEC.csv, or similar.

prefixstr

A prefix to attach to each token in expression.

tokenslist-like of str

A list of tokens to edit within an expression.

Returns
expressionstr

The modified expression

activitysim.estimation.larch.cdap.cdap_base_utility_by_person(model, n_persons, spec, alts=None, value_tokens=())

Build the base utility by person for each pattern.

Parameters
modellarch.Model
n_personsint
specpandas.DataFrame

The base utility by person spec provided by the ActivitySim framework.

altsdict, optional

The keys are the names of the patterns, and the values are the alternative code numbers, as created by generate_alternatives. If not given, the alts are automatically regenerated using that function.

value_tokenslist-like of str, optional

A list of tokens to edit within an the expressions, generally the column names of the provided values from the estimation data bundle. Only used when n_persons is more than 1.

activitysim.estimation.larch.cdap.generate_alternatives(n_persons)

Generate a dictionary of CDAP alternatives.

The keys are the names of the patterns, and the values are the alternative code numbers.

Parameters
n_personsint
Returns
Dict
activitysim.estimation.larch.cdap.interact_pattern(n_persons, select_persons, tag)

Compile a regex pattern to match CDAP alternatives.

Parameters
n_personsint
select_personslist-like of int

The persons to be selected.

tagstr

The activity letter, currently one of {M,N,H}.

Returns
re.compile
activitysim.estimation.larch.scheduling.construct_availability_ca(model, chooser_data, alt_codes_to_names)

Construct an availability dataframe based on -999 parameters.

Parameters
modellarch.Model
chooser_datapandas.DataFrame
alt_codes_to_namesMapping[int,str]
Returns
pandas.DataFrame