assisted.py

"""File assisted.py

:author: Michel Bierlaire
:date: Sun Mar 19 17:06:29 2023

New version of the assisted specification using Catalogs

"""

import logging
from typing import Callable

from biogeme_optimization.neighborhood import Neighborhood, Operator as VnsOperator
from biogeme_optimization.pareto import Pareto, SetElement, DATE_TIME_STRING
from biogeme_optimization.vns import vns, ParetoClass
from matplotlib.axes import Axes

import biogeme.tools.unique_ids
import biogeme.version as bv
from biogeme.biogeme import BIOGEME
from biogeme.configuration import Configuration
from biogeme.controller import ControllerOperator
from biogeme.exceptions import BiogemeError
from biogeme.parameters import Parameters
from biogeme.results import bioResults
from biogeme.specification import Specification

logger = logging.getLogger(__name__)


# Operators


class ParetoPostProcessing:
    """Class to process an existing Pareto set."""

    def __init__(
        self,
        biogeme_object: BIOGEME,
        pareto_file_name: str,
    ):
        """Ctor

        :param biogeme_object: object containing the loglikelihood and the database
        :type biogeme_object: biogeme.biogeme.BIOGEME

        :param pareto_file_name: file where to read and write the Pareto solutions
        :type pareto_file_name: str

        """
        self.biogeme_object = biogeme_object
        self.pareto = Pareto(filename=pareto_file_name)
        self.expression = biogeme_object.log_like
        if self.expression is None:
            error_msg = 'No log likelihood function is defined'
            raise BiogemeError(error_msg)
        self.database = biogeme_object.database
        self.model_names = None

    def reestimate(self, recycle: bool = False) -> dict[str, bioResults]:
        """The assisted specification uses quickEstimate to estimate
        the models. A complete estimation is necessary to obtain the
        full estimation results.

        """
        if self.model_names is None:
            self.model_names = biogeme.tools.unique_ids.ModelNames(
                prefix=self.biogeme_object.modelName
            )

        all_results = {}
        for element in self.pareto.pareto:
            config_id = element.element_id
            the_biogeme = BIOGEME.from_configuration(
                config_id=config_id,
                expression=self.expression,
                database=self.database,
                parameter_file=self.biogeme_object.parameter_file,
            )
            _ = Configuration.from_string(config_id)
            the_biogeme.modelName = self.model_names(config_id)
            the_result = the_biogeme.estimate(recycle=recycle)
            all_results[config_id] = the_result
        return all_results

    def log_statistics(self) -> None:
        """Report some statistics about the process in the logger"""
        for msg in self.pareto.statistics():
            logger.info(msg)

    def plot(
        self,
        objective_x: int = 0,
        objective_y: int = 1,
        label_x: str | None = None,
        label_y: str | None = None,
        margin_x: int = 5,
        margin_y: int = 5,
        ax: Axes | None = None,
    ):
        """Plot the members of the set according to two
            objective functions.  They  determine the x- and
            y-coordinate of the plot.

        :param objective_x: index of the objective function to use for the x-coordinate.
        :type objective_x: int

        :param objective_y: index of the objective function to use for the y-coordinate.
        :type objective_y: int

        :param label_x: label for the x_axis
        :type label_x: str

        :param label_y: label for the y_axis
        :type label_y: str

        :param margin_x: margin for the x axis
        :type margin_x: int

        :param margin_y: margin for the y axis
        :type margin_y: int

        :param ax: matplotlib axis for the plot
        :type ax: matplotlib.Axes

        """

        return self.pareto.plot(
            objective_x, objective_y, label_x, label_y, margin_x, margin_y, ax
        )


class AssistedSpecification(Neighborhood):
    """Class defining assisted specification problem for the VNS algorithm."""

    def __init__(
        self,
        biogeme_object: BIOGEME,
        multi_objectives: Callable[[bioResults], list[float]],
        pareto_file_name: str,
        validity: Callable[[bioResults], bool] | None = None,
        parameter_file: str | None = None,
    ):
        """Ctor

        :param biogeme_object: object containing the loglikelihood and the database
        :type biogeme_object: biogeme.biogeme.BIOGEME

        :param multi_objectives: function calculating the objectives to minimize
        :type multi_objectives: fct(biogeme.results.bioResults) --> list[float]

        :param pareto_file_name: file where to read and write the Pareto solutions
        :type pareto_file_name: str

        :param validity: function verifying that the estimation
            results are valid. It must return a bool and an explanation
            why if it is invalid, or None otherwise

        :type validity: fct(biogeme.results.bioResults) --> Validity

        """
        self.biogeme_parameters: Parameters = Parameters()
        self.biogeme_parameters.read_file(parameter_file)
        self.parameter_file: str = self.biogeme_parameters.file_name
        self.multi_objectives = multi_objectives
        logger.debug('Ctor assisted specification')
        self.biogeme_object = biogeme_object
        self.central_controller = self.biogeme_object.log_like.set_central_controller()
        Specification.generic_name = biogeme_object.modelName
        Specification.user_defined_validity_check = (
            None if validity is None else staticmethod(validity)
        )
        largest_neighborhood = self.biogeme_parameters.get_value(
            name='largest_neighborhood', section='AssistedSpecification'
        )
        self.pareto = ParetoClass(
            max_neighborhood=largest_neighborhood, pareto_file=pareto_file_name
        )
        self.pareto.comments = [
            f'Biogeme {bv.get_version()} [{bv.versionDate}]',
            f'File {self.pareto.filename} created on {DATE_TIME_STRING}',
            f'{bv.AUTHOR}, {bv.DEPARTMENT}, {bv.UNIVERSITY}',
        ]
        self.expression = biogeme_object.log_like
        if self.expression is None:
            error_msg = 'No log likelihood function is defined'
            raise BiogemeError(error_msg)
        self.database = biogeme_object.database
        Specification.expression = self.expression
        Specification.database = self.database
        self.operators = {
            name: self.generate_operator(operator)
            for name, operator in self.central_controller.prepare_operators().items()
        }
        super().__init__(self.operators)

    def generate_operator(self, function: ControllerOperator) -> VnsOperator:
        """Defines an operator that takes a SetElement as an argument, to
            comply with the interface of the VNS algorithm.

        :param function: one of the function implementing the
            operators from the central controller
        :type function: function(str, int) --> str, int

        :return: operator
        :rtype: function(SetElement, int) --> SetElement, int

        """

        def the_operator(element: SetElement, step: int) -> tuple[SetElement, int]:
            the_new_configuration, number_of_modifications = function(
                Configuration.from_string(element.element_id),
                step,
            )
            new_specification = Specification(configuration=the_new_configuration)
            return (
                new_specification.get_element(self.multi_objectives),
                number_of_modifications,
            )

        return the_operator

    def is_valid(self, element: SetElement) -> tuple[bool, str]:
        """Check the validity of the solution.

        :param element: solution to be checked
        :type element: :class:`biogeme.pareto.SetElement`

        :return: valid, why where valid is True if the solution is
            valid, and False otherwise. why contains an explanation why it
            is invalid.
        :rtype: tuple(bool, str)
        """
        if not isinstance(element, SetElement):
            raise BiogemeError(f'Wrong type {type(element)} instead of SetElement')

        specification = Specification.from_string_id(element.element_id)
        return specification.validity

    def run(self) -> dict[str, bioResults]:
        """Runs the VNS algorithm

        :return: doct with the estimation results of the Pareto optimal models
        :rtype: dict[biogeme.results.bioResults]

        """
        logger.debug('Run assisted specification')
        logger.debug('Pareto solutions BEFORE')
        for elem in self.pareto.pareto:
            logger.debug(elem.element_id)
        # We first try to estimate all possible configurations
        Specification.database = self.biogeme_object.database
        Specification.expression = self.biogeme_object.log_like
        Specification.pareto = self.pareto
        logger.debug('Default specification')
        default_specification = Specification.default_specification()
        the_element = default_specification.get_element(self.multi_objectives)
        Specification.pareto.add(the_element)

        logger.info(f'{default_specification=}')
        logger.debug('Default specification: done')
        pareto_before = self.pareto.length_of_all_sets()

        # Check if we can estimate everything
        number_of_specifications = (
            self.biogeme_object.log_like.number_of_multiple_expressions()
        )
        maximum_number = self.biogeme_object.maximum_number_catalog_expressions
        if number_of_specifications <= maximum_number:
            logger.info('We consider all possible combinations of the catalogs.')
            for index, configuration in enumerate(self.biogeme_object.log_like):
                logger.info(f'Model {index}/{number_of_specifications}')
                the_config = configuration.current_configuration()
                the_specification = Specification(the_config)
                the_element = the_specification.get_element(self.multi_objectives)
                Specification.pareto.add(the_element)
                Specification.pareto.dump()
        else:
            logger.info(
                f'The number of possible specifications [{number_of_specifications}] '
                f'exceeds the maximum number [{maximum_number}]. '
                f'A heuristic algorithm is applied.'
            )

            default_element = default_specification.get_element(self.multi_objectives)
            number_of_neighbors = self.biogeme_parameters.get_value(
                name='number_of_neighbors', section='AssistedSpecification'
            )
            maximum_attempts = self.biogeme_parameters.get_value(
                name='maximum_attempts', section='AssistedSpecification'
            )
            logger.debug(f'{default_element=}')
            self.pareto = vns(
                problem=self,
                first_solutions=[default_element],
                pareto=self.pareto,
                number_of_neighbors=number_of_neighbors,
                maximum_attempts=maximum_attempts,
            )
        logger.debug('Pareto solutions AFTER')
        for elem in self.pareto.pareto:
            logger.debug(elem.element_id)

        pareto_after = self.pareto.length_of_all_sets()

        self.pareto.dump()
        logger.info(f'Pareto file has been updated: {self.pareto.filename}')
        logger.info(
            f'Before the algorithm: {pareto_before[1]} models, '
            f'with {pareto_before[0]} Pareto.'
        )
        logger.info(
            f'After the algorithm: {pareto_after[1]} models, '
            f'with {pareto_after[0]} Pareto.'
        )

        # Postprocessing
        logger.info(
            'VNS algorithm completed. Postprocessing of the Pareto optimal solutions'
        )

        post_processing = ParetoPostProcessing(
            biogeme_object=self.biogeme_object, pareto_file_name=self.pareto.filename
        )
        estimation_results = post_processing.reestimate()
        post_processing.log_statistics()
        return estimation_results