rssalessio / nnGA

Neural Network Genetic Algorithm library used for deep learning problems

nnGA Library - Neural Network Genetic Algorithm Library (v0.3)

Off the shelf Genetic Algorithm library for deep learning problems

Author: Alessio Russo (PhD Student at KTH - [e-mail])

License

Our code is released under the MIT license (refer to the LICENSE file for details).

Requirements

To use the library you need atleast Python 3.5. Examples may require additional libraries.

Other required dependencies:

• NumPy

Usage/Examples

You can import the library by typing python from nnga import nnGA.

To learn how to use nnGA, check the following examples:

1. Reinforcement learning examples

   • Cartpole example
   • Lunarlander example

2. Supervised learning examples

   • Multiclass classification

In general the code has the following structure

from nnga import nnGA, GaussianInitializationStrategy, \
    GaussianMutationStrategy, BasicCrossoverStrategy, \
    PopulationParameters

def make_network(parameters=None):
    ''' Function that creates a network given a set of parameters '''
    neural_network = ...
    return neural_network


def fitness(idx, parameters):
    ''' Fitness function to evaluate a set of parameters '''
    # Evaluate parameters
    network = make_network(parameters)
    return evaluate_network(network)


if __name__ == '__main__':
    # Initialize GA parameters
    network = make_initial_network()
    network_structure = [list(layer.shape) for layer in network]  # List of tuples, containing the shape of each layer
    
    # Population parameters
    population = PopulationParameters(population_size=200)
    
    # Mutation strategy
    mutation = GaussianMutationStrategy(network_structure, 1e-1)
    
    # Crossover strategy
    crossover = BasicCrossoverStrategy(network_structure)
    
    # Initialization strategy
    init = GaussianInitializationStrategy(
        mean=0., std=1., network_structure=network_structure)

    ga = nnGA(
        epochs=50,  # Number of epochs
        fitness_function=fitness,
        population_parameters=population,
        mutation_strategy=mutation,
        initialization_strategy=init,
        crossover_strategy=crossover,
        num_processors=8)  # Number of cores

    # Run GA
    network_parameters, best_result, results = ga.run()