Creating surrogate models in Python with EvoNN, EvoDN2 and BioGP¶
+This example will show how to use the code and structure in pyRVEA to create models using EvoNN, EvoDN2 and BioGP (see descriptions of the algorithms here). The code is currently capable of training and optimizing the models with all of the genetic algorithms implemented in pyRVEA/EAs.
+The basic workflow is as follows:
+-
+
Create or import training data
+Create a problem class which handles the training of the models
+After training, create a new population and evolve it using the models to optimize the problem
+
For training data, example functions can be found in pyRVEA/Problem/testproblem.py and test_functions.py packages, or a custom data set can be imported.
+Using Fonseca-Fleming two objective function with 2 variables as an example:
+import numpy as np
+import pandas as pd
+from pyrvea.Problem.testproblem import TestProblem
+from pyrvea.Problem.dataproblem import DataProblem
+
+test_prob = TestProblem(name="Fonseca-Fleming", num_of_variables=2)
+
+# Random data for training
+# x = list of variable names, y = list of objectives
+dataset, x, y = test_prob.create_training_data(samples=500)
+Or you can import a data set, here an example data set for ZDT1 problem with 30 variables and 1000 samples is used. The DataProblem class requires the names of the variables and objectives as lists. If your data set does not contain them in the header, you will have to create them yourself.
+dataset = pd.read_excel("ZDT1_1000.xls", header=0)
+x = dataset.columns[0:30].tolist()
+# x = ['x1', 'x2', 'x3', ...]
+y = dataset.columns[30:].tolist()
+# y = ['f1', 'f2']
+After you have the data, create the DataProblem class and pass the data, variables and objectives.
+problem = DataProblem(data=dataset, x=x, y=y)
+You can split the data into a training and testing set:
+problem.train_test_split(train_size=0.7)
+Parameters can be passed for the model and for the genetic algorithm as dictionaries. If no parameters are passed, the defaults are used. +Select the model and the genetic algorithm you want to use and set the parameters (or use defaults). Check EA documentation for details for the genetic algorithms and see docs for EvoNN parameters, EvoDN2 parameters and BioGP parameters.
+Both EA parameters and the model parameters can greatly affect the performance of the model. The best options depend on the problem, so experimenting with different values is encouraged.
+from pyrvea.EAs.PPGA import PPGA
+
+ea_parameters = {
+ "generations_per_iteration": 10,
+ "iterations": 10,
+}
+
+model_parameters = {
+ "training_algorithm": PPGA
+}
+
+problem.train(
+ model_type="EvoDN2",
+ model_parameters=model_parameters,
+ ea_parameters=ea_parameters)
+Note that for BioGP, function set and terminal should be adjusted according to the problem. By default, function set includes addition, substraction, multiplication and division, and terminal set includes the variables from the data. +For Fonseca-Fleming function, square root and negative should be added to the function set and constants 1 and 2 to the terminal set:
+f_set = ("add", "sub", "mul", "div", "sqrt", "neg")
+t_set = [1, 2]
+
+model_parameters = {
+ function_set = f_set
+ terminal_set = t_set
+}
+
+problem.train(
+ model_type="BioGP",
+ model_parameters=model_parameters
+)
+The models can currently be trained with available algorithms in pyRVEA/EAs. For explanations of the different EAs, see their respective documentation.
+The model’s prediction vs. target values can be plotted as follows:
+# Prediction for f1
+f1_y_pred = problem.models["f1"][0].predict(problem.data[problem.x])
+
+problem.models["f1"][0].plot(
+ f1_y_pred, np.asarray(problem.data["f1"]), name="filename" + "f1"
+)
+# Prediction for f2
+f2_y_pred = problem.models["f2"][0].predict(problem.data[problem.x])
+
+problem.models["f2"][0].plot(
+ f2_y_pred, np.asarray(problem.data["f2"]), name="filename" + "f2"
+)
+After the models have been trained, the objectives can be optimized by creating a new population, passing the data problem class (containing the trained models) and calling evolve. EA parameters can be modified for optimization phase if wanted.
+from pyrvea.EAs.RVEA import RVEA
+
+pop = Population(problem)
+opt_params = {"iterations": 10, "generations_per_iteration": 25}
+pop.evolve(EA=RVEA, ea_parameters=opt_params)
+To show the final pareto plot:
+pop.plot_pareto(name="my-test-function")
+