darwin-mendel 0.1.5

Genetic Algorithm: Optimize the output of machine learning models

Genetic Algorithm: A unique way for hyper-parameter tuning of ML models.

The process of evolution and natural selection used in this project to select the best hyper-parameters for certain regression techniques like Decision Tree Regression, Random Forest Regression, Light Gradient Boosting Regression and Extreme Gradient Boosting Regression.

1: Install the library

pip install darwin-mendel

2: Example for Random Forest Regression

from darwin_mendel.optimize_rfr import optimize_rfr
import sklearn.datasets as datasets

iris = datasets.load_boston()
df = pd.DataFrame(iris.data)
x_train, x_test, y_train, y_test = train_test_split(df[[2,4,5,6,7,8,9,10,11]], 
                                                    df[12], test_size=0.2, random_state=2021)

model, hyp_param = optimize_xgbr(x_train=x_train, y_train=y_train, y_test=y_test, x_test=x_test,
                           number_of_generation=10, population_size=30, error_metric='RMSE', mutation_rate=0.1)
print(hyp_param)

3: OutPut:

    n_estimators        368
    max_depth             9
    learning_rate       0.2
    booster          gbtree
    reg_alpha             0
    reg_lambda            1
    RMSE              16.76
    Name: 0, dtype: object

4. Arguments:

    a. User must provide the x_train, y_train, x_test and y_test in the arguments. They should not 
       contain any missing values and strings values.
    b. User can select the error_metreic between "MAPE" and "RMSE", it is used to select the best model.
       Default is "MAPE" (Mean Absolute Percentage Error).
    c. population_size defines initial number of combination of hyper-parameters from which off-springs
       are produced. Thumb rule is, it should be: 5 * number of variables. Default is 50.
    d. number_of_generation is the number of new batches produced from the initial population.
       The more the number of generation, the better will be the resut but it could increase the 
       time consumption. Default is 10.
    e. mutation_rate is the percentage impurity added in a new batch of off springs, it helps in
       reaching the global minimum. Default is 0.05 i.e. 5%
    f. random_seed has to be fixed to make the results repeatable. Default is 2021.
    g. params: {"n_estimators": [2,3,4,.....1000],
                "max_features": ['sqrt', 'auto', 'log2', None],
                "min_samples_leaf": [2,3,4,5,6,.....16],
                "max_depth": [2,3,4,5,6,.....20]}
       Above mentioned are the default ranges for each hyperparameter of Random Forest Regression.
       User can give the range according to her need.
   

5. Default params ranges for other algorithms.

    a. XGBR: 
            {"n_estimators": [2,3,4,5,6,......1000],
             "max_depth": [2,3,4,5,6....20],
             "learning_rate": [0.001, 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1],
             "booster": ["gbtree"],
             "reg_alpha": [0],
             "reg_lambda": [1]}  
    b. LGBMR: 
            {"n_estimators": [2,3,4,5,6,......1000],
             "max_depth": [2,3,4,5,6....20],
             "learning_rate": [0.001, 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1],
             "boosting_type": ["gbdt"],
             "num_leaves": [2,3,4,5,6,....15],
             "reg_alpha": [0],
             "reg_lambda": [0]}  
    c. DTR: 
            {"min_samples_leaf": [1,2,3,4,5,6....20],
             "max_depth": [2,3,4,5,6....20],
             "max_features": ["auto", "sqrt", "log2"],
             "splitter": ["best", "random"],
             "criterion": ["mse", "friedman_mse", "mae"]}