Deap Hypervolume comparison for varying population size and number of objectives, testing for use with high dimensional NSGA III with many objectives (over 8).) Discussion here: https://groups.google.com/g/deap-users/c/XLfLa3at6pw

hypervol_man_dim_test.py

# Derek Tishler

# Deap Hypervolume comparison for varying population size and number of objectives, 
# testing for use with high dimensional NSGA III with many objectives (over 8).)

# Discussion here, also required fitness_values.txt with 212x14 array of fitness values per ind in pop:
# https://groups.google.com/g/deap-users/c/XLfLa3at6pw

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from time import time

from deap.tools._hypervolume import hv

hv_calc_max = 10. #seconds max we allow the hv to compute before stopping scan

def perform_stress_test(pop_fits=None, ref_hyp_vals=None, hypervolume_f=None, label=None):
    results = {}
    results['ndim'] = []
    results['hypervolume'] = []
    results['elapsed'] = []
     
    for i in range(1, len(pop_fits)):
        start_time = time()
        value_hypervolume = hypervolume_f(pop_fits[:i], ref_hyp_vals)
        elapsed = time()-start_time
        print('n_obj:%d n_pop:%d / %d\thypervolume: %g\telapsed: %g sec' % (pop_fits.shape[1], 
                                                                            i, 
                                                                            len(pop_fits), 
                                                                            value_hypervolume, 
                                                                            elapsed))
        results['ndim'].append(i)
        results['hypervolume'].append(value_hypervolume)
        results['elapsed'].append(elapsed)

        # Give up once too slow (>20 sec per hv calc)
        if elapsed > hv_calc_max:
            break

    
    result_df = pd.DataFrame.from_dict(results)
    
    plt.figure(figsize=(6,4), dpi=150)
    plt.plot(result_df['ndim'].values, result_df['elapsed'].values)
    plt.title(label)
    plt.xlabel('n_pop / %d' %len(ttt))
    plt.ylabel('hypervolume comp time (seconds)')
    plt.grid()
    plt.tight_layout()
    plt.savefig(label+'.png')
    plt.clf()
    plt.close()

    return result_df


ttt = np.loadtxt('fitness_values.txt')

ref_hyp = np.array([19.685825,962.5723419,5.69326178,0.043162976,
                    11.61159696,0.01699155,0.018175168,223965.1889,
                    2.61386E-06,0.016530917,1.9237E-06,2.82397E-06,
                    4472.93138,4.81537E-06])


if __name__ == "__main__":
    ### 64 bit
    print('Fitness dtype(pop & ref):')
    print(ttt.dtype)
    print(ref_hyp.dtype)

    n_obj_1 = 7
    test_1_df = perform_stress_test(pop_fits = ttt[:,:n_obj_1], 
                        ref_hyp_vals = ref_hyp[:n_obj_1], 
                        hypervolume_f = hv.hypervolume, 
                        label = '%d-Objective'%n_obj_1)

    n_obj_2 = 10
    test_2_df = perform_stress_test(pop_fits = ttt[:,:n_obj_2], 
                        ref_hyp_vals = ref_hyp[:n_obj_2], 
                        hypervolume_f = hv.hypervolume, 
                        label = '%d-Objective'%n_obj_2)

    test_3_df = perform_stress_test(pop_fits = ttt, 
                        ref_hyp_vals = ref_hyp, 
                        hypervolume_f = hv.hypervolume, 
                        label = '%d-Objective'%len(ref_hyp))


    plt.figure(figsize=(6,4), dpi=150)
    plt.semilogy(test_1_df['ndim'].values, test_1_df['elapsed'].values, label='%d Obj'%n_obj_1, c='dodgerblue')
    plt.semilogy(test_2_df['ndim'].values, test_2_df['elapsed'].values, label='%d Obj'%n_obj_2, c='forestgreen')
    plt.semilogy(test_3_df['ndim'].values, test_3_df['elapsed'].values, label='%d Obj'%len(ref_hyp), c='orangered')
    plt.legend()
    plt.axhline(hv_calc_max, c='r', ls='--')
    plt.title("deap hypervolume")
    plt.xlabel('n_pop / %d' %len(ttt))
    plt.ylabel('hypervolume comp time (seconds)')
    plt.grid()
    plt.tight_layout()
    plt.savefig('hv_summary.png')
    plt.clf()
    plt.close()

Output Plots: