A self organizing map algorithm for scheduling meeting times as availabilities and bookings.  A map is a low-dimensional representation of a training sample comprising of elements e. It is represented by nodes n. The map is transformed by a  regression operation to modify the nodes position one element from the model (e) at a time. With preferences translating to nodes and availabilities as elements, this allows the map to start getting a closer match to the sample space with each epoch/iteration.

from sys import argv

import numpy as np

from io_helper import read_xyz, normalize

from neuron import generate_network, get_neighborhood, get_boundary

from distance import select_closest, euclidean_distance, boundary_distance

from plot import plot_network, plot_boundary

def main():

    if len(argv) != 2:

        print("Correct use: python src/main.py <filename>.xyz")

        return -1

    problem = read_xyz(argv[1])

    boundary = som(problem, 100000)

    problem = problem.reindex(boundary)

    distance = boundary_distance(problem)

    print('Boundary found of length {}'.format(distance))

def som(problem, iterations, learning_rate=0.8):

    """Solve the xyz using a Self-Organizing Map."""

    # Obtain the normalized set of timeslots (w/ coord in [0,1])

    timeslots = problem.copy()

    # print(timeslots)

    #timeslots[['X', 'Y', 'Z']] = normalize(timeslots[['X', 'Y', 'Z']])

    # The population size is 8 times the number of timeslots

    n = timeslots.shape[0] * 8

    # Generate an adequate network of neurons:

    network = generate_network(n)

    print('Network of {} neurons created. Starting the iterations:'.format(n))

    for i in range(iterations):

        if not i % 100:

            print('\t> Iteration {}/{}'.format(i, iterations), end="\r")

        # Choose a random timeslot

        timeslot = timeslots.sample(1)[['X', 'Y', 'Z']].values

        winner_idx = select_closest(network, timeslot)

        # Generate a filter that applies changes to the winner's gaussian

        gaussian = get_neighborhood(winner_idx, n//10, network.shape[0])

        # Update the network's weights (closer to the timeslot)

        network += gaussian[:,np.newaxis] * learning_rate * (timeslot - network)

        # Decay the variables

        learning_rate = learning_rate * 0.99997

        n = n * 0.9997

        # Check for plotting interval

        if not i % 1000:

            plot_network(timeslots, network, name='diagrams/{:05d}.png'.format(i))

        # Check if any parameter has completely decayed.

        if n < 1:

            print('Radius has completely decayed, finishing execution',

            'at {} iterations'.format(i))

            break

        if learning_rate < 0.001:

            print('Learning rate has completely decayed, finishing execution',

            'at {} iterations'.format(i))

            break

    else:

        print('Completed {} iterations.'.format(iterations))

    # plot_network(timeslots, network, name='diagrams/final.png')

    boundary = get_boundary(timeslots, network)

    plot_boundary(timeslots, boundary, 'diagrams/boundary.png')

    return boundary

if __name__ == '__main__':

    main()

Reference: 

https://github.com/raja0034/som4drones

DroneCommercialSoftware.docx