Optga: multi-objective optimizer for ML-model outputs

Finding out pareto-optimum inputs of your machine learning models in order to balance multipule objective variables which in trade-off relationship

Overview

In real world, machine learning projects dealing with scientific issues often have two or more target variables which in trade-off relationship.

For example:

• Speed and fuel efficiency in automobiles.
• Lower voltage and higher capacity in Li-ion batteries.
• Wing weight and lift in aircrafts
• Permeability and selectivity in polymer membranes

After you have successfully built predictive models for each target variable, Optga finds out pareto-optimum inputs of your machine learning models in order to help making a dicision and achieve desired outcomes.

Key features

• Rapid multi-objective optimization by genetic algorithm (NSGA2) with numba
• Support various constraints specific to machine learning inputs. (e.g. Onehot constraint, descrete constraint)
• Easy user interface

Installation

If you are using conda, updating numba may be required before installation.

pip install optga

Getting started

Exploring the trade-off between house prices and age in Boston dataset.

import pandas as pd
from sklearn.datasets import load_boston
from sklearn.ensemble import RandomForestRegressor
import optga
from optga.optimizer import Optimizer

#: Load boston housing price datasets
df =  pd.DataFrame(load_boston().data,
                   columns=load_boston().feature_names)
X = df.drop(["AGE"], 1)
y_price = pd.DataFrame(boston.target, columns=["Price"])
y_age = pd.DataFrame(df["AGE"])

#: create predictive models
model_price = RandomForestRegressor().fit(X, y_price)
model_age = RandomForestRegressor().fit(X, y_age)

optimizer = Optimizer(sample_data=X)

#: Add scoring function
optimizer.add_objective("Price", model_price.predict, direction="minimize")
optimizer.add_objective("Age", model_age.predict, direction="minimize")

#: Add constraints on explanatory variables
optimizer.add_discrete_constraint("CHAS", [0, 1])
optimizer.add_discrete_constraint("ZN", [0, 100])
optimizer.add_discrete_constraint("RAD", list(range(1, 9)) + [24])

#: Run optimization and export results
optimizer.run(n_gen=300, population_size=300)
optimizer.export_result("boston_result")

See full examples on example/1_BostonHousingPrice.ipynb.

ToDo

• Add island strategy as multi-processing GA optimization.