Source code for l2l.optimizers.face.optimizer

import logging
from collections import namedtuple

import numpy as np

from l2l import dict_to_list, list_to_dict
from l2l.optimizers.optimizer import Optimizer

logger = logging.getLogger("optimizers.face")

FACEParameters = namedtuple('FACEParameters',
                            ['min_pop_size', 'max_pop_size', 'n_elite', 'smoothing', 'temp_decay', 'n_iteration',
                             'distribution', 'stop_criterion', 'n_expand', 'seed'])

FACEParameters.__doc__ = """
:param min_pop_size: Minimal number of individuals per simulation.
:param max_pop_size: This is the minimum amount of samples taken into account for the FACE algorithm
:param n_elite: Number of individuals to be considered as elite
:param smoothing: This is a factor between 0 and 1 that determines the weight assigned to the previous distribution
  parameters while calculating the new distribution parameters. The smoothing is done as a linear combination of the 
  optimal parameters for the current data, and the previous distribution as follows:
    
    new_params = smoothing*old_params + (1-smoothing)*optimal_new_params
    
:param temp_decay: This parameter is the factor (necessarily between 0 and 1) by which the temperature decays each
  generation. To see the use of temperature, look at the documentation of :class:`.FACEOptimizer`
:param n_iteration: Number of iterations to perform
:param distribution: Distribution class to use. Has to implement a fit and sample function.
:param stop_criterion: (Optional) Stop if this fitness is reached.
:param n_expand: (Optional) This is the amount by which the sample size is increased if FACE becomes active
"""


[docs]class FACEOptimizer(Optimizer): """ Class for Fully Adaptive Crossentropy Optimizer (adaptive sample size) In the pseudo code the algorithm does: For n iterations do: 1. Sample individuals from distribution 2. evaluate individuals and get fitness 3. check if gamma or best individuals fitness increased 4. if not increase population size by n_expand (if not yet max_pop_size else stop) and sample again (1) else set pop_size = min_pop_size and proceed 5. pick n_elite individuals with highest fitness 6. Out of the remaining non-elite individuals, select them using a simulated-annealing style selection based on the difference between their fitness and the `1-rho` quantile (*gamma*) fitness, and the current temperature 7. Fit the distribution family to the new elite individuals by minimizing cross entropy. The distribution fitting is smoothed to prevent premature convergence to local minima. A weight equal to the `smoothing` parameter is assigned to the previous parameters when smoothing. return final distribution parameters. (The final distribution parameters contain information regarding the location of the maxima) :param ~l2l.utils.trajectory.Trajectory traj: Use this trajectory to store the parameters of the specific runs. The parameters should be initialized based on the values in `parameters` :param optimizee_create_individual: Function that creates a new individual. All parameters of the Individual-Dict returned should be of numpy.float64 type :param optimizee_fitness_weights: Fitness weights. The fitness returned by the Optimizee is multiplied by these values (one for each element of the fitness vector) :param parameters: Instance of :func:`~collections.namedtuple` :class:`.FACEParameters` containing the parameters needed by the Optimizer """ def __init__(self, traj, optimizee_create_individual, optimizee_fitness_weights, parameters, optimizee_bounding_func=None): super().__init__(traj, optimizee_create_individual=optimizee_create_individual, optimizee_fitness_weights=optimizee_fitness_weights, parameters=parameters, optimizee_bounding_func=optimizee_bounding_func) self.optimizee_bounding_func = optimizee_bounding_func if parameters.min_pop_size < 1: raise Exception("min_pop_size needs to be greater than 0") if parameters.max_pop_size < parameters.min_pop_size: raise Exception("max_pop_size needs to be greater or equal to min_pop_size") if parameters.n_elite > parameters.min_pop_size: raise Exception("n_elite exceeds min_pop_size") if parameters.temp_decay < 0 or parameters.temp_decay > 1: raise Exception("temp_decay not in range") if parameters.smoothing >= 1 or parameters.smoothing < 0: raise Exception("smoothing has to be in interval [0, 1)") if parameters.seed is None: raise Exception("The 'seed' must be set") # The following parameters are recorded traj.f_add_parameter('min_pop_size', parameters.min_pop_size, comment='Number of minimal individuals simulated in each run') traj.f_add_parameter('max_pop_size', parameters.max_pop_size, comment='Maximal individuals in population') traj.f_add_parameter('n_elite', parameters.n_elite, comment='Number of individuals to be considered as elite') traj.f_add_parameter('n_iteration', parameters.n_iteration, comment='Number of iterations to run') traj.f_add_parameter('n_expand', parameters.n_expand, comment='Expanding of population size in case of FACE') traj.f_add_parameter('stop_criterion', parameters.stop_criterion, comment='Stop if best individual reaches this fitness') traj.f_add_parameter('smoothing', parameters.smoothing, comment='Weight of old parameters in smoothing') traj.f_add_parameter('temp_decay', parameters.temp_decay, comment='Decay factor for temperature') traj.f_add_parameter('seed', np.uint32(parameters.seed), comment='Random seed used by optimizer') self.random_state = np.random.RandomState(seed=traj.par.seed) temp_indiv, self.optimizee_individual_dict_spec = dict_to_list(self.optimizee_create_individual(), get_dict_spec=True) traj.f_add_derived_parameter('dimension', len(temp_indiv), comment='The dimension of the parameter space of the optimizee') # Added a generation-wise parameter logging traj.results.f_add_result_group('generation_params', comment='This contains the optimizer parameters that are' ' common across a generation') # The following parameters are recorded as generation parameters i.e. once per generation self.g = 0 # the current generation # This is the value above which the samples are considered elite in the # current generation self.gamma = -np.inf self.T = 1 # This is the temperature used to filter evaluated samples in this run self.pop_size = parameters.min_pop_size # Population size is dynamic in FACE self.best_fitness_in_run = -np.inf self.best_individual = None # The first iteration does not pick the values out of the Gaussian distribution. It picks randomly # (or at-least as randomly as optimizee_create_individual creates individuals) # Note that this array stores individuals as an np.array of floats as opposed to Individual-Dicts # This is because this array is used within the context of the cross entropy algorithm and # Thus needs to handle the optimizee individuals as vectors current_eval_pop = [self.optimizee_create_individual() for _ in range(parameters.min_pop_size)] if optimizee_bounding_func is not None: current_eval_pop = [self.optimizee_bounding_func(ind) for ind in current_eval_pop] self.eval_pop = current_eval_pop self.eval_pop_asarray = np.array([dict_to_list(x) for x in self.eval_pop]) # Max Likelihood self.current_distribution = parameters.distribution self.current_distribution.init_random_state(self.random_state) self.current_distribution.fit(self.eval_pop_asarray) self._expand_trajectory(traj)
[docs] def post_process(self, traj, fitnesses_results): """ See :meth:`~l2l.optimizers.optimizer.Optimizer.post_process` """ n_elite, n_iteration, smoothing, temp_decay, min_pop_size, max_pop_size = \ traj.n_elite, traj.n_iteration, traj.smoothing, traj.temp_decay, traj.min_pop_size, traj.max_pop_size stop_criterion, n_expand = traj.stop_criterion, traj.n_expand weighted_fitness_list = [] # ************************************************************************************************************** # Storing run-information in the trajectory # Reading fitnesses and performing distribution update # ************************************************************************************************************** for run_index, fitness in fitnesses_results: # We need to convert the current run index into an ind_idx # (index of individual within one generation) traj.v_idx = run_index ind_index = traj.par.ind_idx traj.f_add_result('$set.$.individual', self.eval_pop[ind_index]) traj.f_add_result('$set.$.fitness', fitness) weighted_fitness_list.append(np.dot(fitness, self.optimizee_fitness_weights)) traj.v_idx = -1 # set trajectory back to default # Performs descending arg-sort of weighted fitness fitness_sorting_indices = list(reversed(np.argsort(weighted_fitness_list))) generation_name = 'generation_{}'.format(self.g) # Sorting the data according to fitness sorted_population = self.eval_pop_asarray[fitness_sorting_indices] sorted_fitess = np.asarray(weighted_fitness_list)[fitness_sorting_indices] # Elite individuals are with performance better than or equal to the (1-rho) quantile. # See original describtion of cross entropy for optimization elite_individuals = sorted_population[:n_elite] previous_best_fitness = self.best_fitness_in_run self.best_individual = list_to_dict(sorted_population[0], self.optimizee_individual_dict_spec) self.best_fitness_in_run = sorted_fitess[0] previous_gamma = self.gamma self.gamma = sorted_fitess[n_elite - 1] logger.info("-- End of generation %d --", self.g) logger.info(" Evaluated %d individuals", len(fitnesses_results)) logger.info(' Best Fitness: %.4f', self.best_fitness_in_run) logger.debug(' Calculated gamma: %.4f', self.gamma) # ************************************************************************************************************** # Storing Generation Parameters / Results in the trajectory # ************************************************************************************************************** # These entries correspond to the generation that has been simulated prior to this post-processing run traj.results.generation_params.f_add_result_group(generation_name, 'New generation added to results') traj.results.generation_params.f_add_result(generation_name + '.g', self.g, comment='The index of the evaluated generation') traj.results.generation_params.f_add_result(generation_name + '.gamma', self.gamma, comment='The fitness threshold inferred from the evaluated ' 'generation (This is used in sampling the next generation') traj.results.generation_params.f_add_result(generation_name + '.T', self.T, comment='Temperature used to select non-elite elements among the' 'individuals of the evaluated generation') traj.results.generation_params.f_add_result(generation_name + '.best_fitness_in_run', self.best_fitness_in_run, comment='The highest fitness among the individuals in the ' 'evaluated generation') traj.results.generation_params.f_add_result(generation_name + '.pop_size', self.pop_size, comment='Population size') # Check stopping if self.g >= n_iteration or self.best_fitness_in_run >= stop_criterion: return expand = True if self.best_fitness_in_run > previous_best_fitness or self.gamma > previous_gamma: # shrink population size self.pop_size = (self.pop_size + min_pop_size) // 2 # new distribution fit individuals_to_be_fitted = elite_individuals # Temperature dependent sampling of non elite individuals if temp_decay > 0: # Keeping non-elite samples with certain probability dependent on temperature (like Simulated Annealing) non_elite_selection_probs = np.clip(np.exp((weighted_fitness_list[n_elite:] - self.gamma) / self.T), amin=0.0, a_max=1.0) non_elite_selected_indices = self.random_state.binomial(1, p=non_elite_selection_probs) non_elite_eval_pop_asarray = sorted_population[n_elite:][non_elite_selected_indices] individuals_to_be_fitted = np.concatenate((elite_individuals, non_elite_eval_pop_asarray)) # Fitting New distribution parameters. self.distribution_results = self.current_distribution.fit(individuals_to_be_fitted, smoothing) elif self.pop_size + n_expand <= max_pop_size: # Increase pop size by one, resample, FACE part logger.info(' FACE increase population size by %d', n_expand) self.pop_size += n_expand else: # Stop algorithm expand = False logger.warning(' Possibly diverged') # Add the results of the distribution fitting to the trajectory for parameter_key, parameter_value in self.distribution_results.items(): traj.results.generation_params.f_add_result(generation_name + '.' + parameter_key, parameter_value) # ************************************************************************************************************** # Create the next generation by sampling the inferred distribution # ************************************************************************************************************** # Note that this is only done in case the evaluated run is not the last run fitnesses_results.clear() self.eval_pop.clear() if expand: # Sample from the constructed distribution self.eval_pop_asarray = self.current_distribution.sample(self.pop_size) self.eval_pop = [list_to_dict(ind_asarray, self.optimizee_individual_dict_spec) for ind_asarray in self.eval_pop_asarray] # Clip to boundaries if self.optimizee_bounding_func is not None: self.eval_pop = [self.optimizee_bounding_func(individual) for individual in self.eval_pop] self.eval_pop_asarray = np.array([dict_to_list(x) for x in self.eval_pop]) self.g += 1 # Update generation counter self.T *= temp_decay self._expand_trajectory(traj)
[docs] def end(self, traj): """ See :meth:`~l2l.optimizers.optimizer.Optimizer.end` """ best_last_indiv_dict = self.best_individual traj.f_add_result('final_individual', best_last_indiv_dict) traj.f_add_result('final_fitness', self.best_fitness_in_run) traj.f_add_result('n_iteration', self.g) # ------------ Finished all runs and print result --------------- # logger.info("-- End of (successful) FACE optimization --") logger.info("-- Final distribution parameters --") for parameter_key, parameter_value in sorted(self.distribution_results.items()): logger.info(' %s: %s', parameter_key, parameter_value)