Parameter Grid Search

The purpose of the grid search is to provide an easy way to tune parameters for a given algorithm in a way that will work well on a certain class of problem. This is accomplished by creating a grid of parameter permutations based on user input and running a selected algorithm with those parameters to find the best set. The best results of every permutation of parameters are saved for post-processing.

---

class optiseek.modelhelpers.parameter_grid_search(algorithm, input_function, var_list, param_grid, optimize_options, show_progress=False)

---

Parameters

Parameter	Description
algorithm : class	Class of the algorithm that you would like to use.
input_function : function	Function object for the algorithm to optimize.
var_list : list of variables	List of all variable objects to define their names and domains in the search space. This is the same list you would pass to the optimization algorithm class.
param_grid : dict	Dictionary containing the grid of parameters to be explored with the parameter names (strings) as keys and a list of parameter values as values. All permutations of parameters in this dict will be tested. For any parameters not specified, the default will be used. See the example for more details.
optimize_options : dict	Dictionary containing the kwargs for the optimize() method of the algorithm.
show_progress : bool	Boolean to indicate whether the grid search will print progress to the console as the solve continues. The number of permutations increases exponentially with respect to parameter inputs, so for high numbers of parameter inputs, this can be useful to see how much longer the solver has left.

---

Attributes

Attribute	Description
best_parameters : dict	A dictionary containing the best performing set of parameters. The parameter names as strings are stored as keys and the corresponding values are stored as values.
best_position : list or ndarray	The most optimal position that was found using the best performing parameters.
best_value : float	The most optimal function value that was found using the best performing parameters.
results : pd.DataFrame	A pandas DataFrame containing all results from the search. Columns represent the algorithm parameters, best position found, and best function value found with the respective parameters.

---

Methods

.search()

Executes the parameter grid search process and stores the results in the class attributes.

---

Example

from optiseek.variables import var_float
from optiseek.modelhelpers import parameter_grid_search
from optiseek.metaheuristics import particle_swarm_optimizer
from optiseek.testfunctions import ackley

# define the variables for the Ackley2D function
var_list = [
    var_float('x', [-10, 10]),
    var_float('y', [-10, 10])
]

# set up the param_grid dictionary
param_grid = {
    'n_particles': [10],
    'weight': [0.20, 0.35, 0.50],
    'phi_p': [1.0, 1.5, 2.0],
    'phi_g': [1.0, 1.5, 2.0],
    'zero_velocity': [True, False]
}

# set up the optimize_options dictionary
optimize_options = {
    'find_minimum': True,
    'max_function_evals': 75
}

# create the an instance of the grid search class
pgs = parameter_grid_search(particle_swarm_optimizer, 
                            ackley, 
                            var_list, 
                            param_grid, 
                            optimize_options)

# start the search
pgs.search()

# show the optimal parameters, the best function value found, and a preview of all saved results
print(f'best parameters: {pgs.best_parameters}')
print(f'best value: {pgs.best_value}\n')
print(pgs.results.head())

best parameters: {'n_particles': 10, 'weight': 0.2, 'phi_p': 1.5, 'phi_g': 1.0, 'zero_velocity': False}
best value: 0.026961543641856878

   n_particles  weight  phi_p  ...         x         y  best_value
0           10    0.20    1.5  ...  0.007717  0.004236    0.026962
1           10    0.50    1.0  ...  0.021830 -0.003103    0.075259
2           10    0.20    2.0  ...  0.019337 -0.018130    0.093595
3           10    0.35    1.0  ... -0.028123  0.002833    0.101083
4           10    0.20    1.5  ...  0.001063  0.031908    0.117212

[5 rows x 8 columns]