NeuralNetwork::NeuralNetwork(const NeuralNetworkTopology& topology)
: m_topology(topology)
{
unsigned int prev_layer_num_neuron = m_topology.getInput();
m_layerHidden.resize( m_topology.getHiddens().size() );
for (unsigned int i = 0; i < m_layerHidden.size(); ++i)
{
t_layer& layer = m_layerHidden[i];
layer.resize( m_topology.getHiddens()[i] );
for (t_neuron& neuron : layer)
{
neuron.m_weights.resize( prev_layer_num_neuron );
for (float& weight : neuron.m_weights)
weight = randomClamped();
}
prev_layer_num_neuron = m_topology.getHiddens()[i];
}
m_layerOutput.resize( m_topology.getOutput() );
for (t_neuron& neuron : m_layerOutput)
{
neuron.m_weights.resize( prev_layer_num_neuron );
for (float& weight : neuron.m_weights)
weight = randomClamped();
}
}
Neuron::Neuron(int newNumInputs): numInputs(newNumInputs+1){
//Need an additional weight for the bias
for (int i=0; i<newNumInputs+1;i++){
//Set up weights with a random initial value
vecWeight.push_back(randomClamped());
}
}
void GeneticAlgorithm::mutate(t_genome& genome) const
{
int total_mutated = 0;
for (float& weight : genome.m_weights)
if (randomFloat() < 0.1f)
{
weight += (randomClamped() * 0.2f);
++total_mutated;
}
// D_MYLOG("total_mutated=" << ((float)total_mutated / genome.m_weights.size()) * 100);
}
void GeneticAlgorithm::BreedPopulation()
{
// D_MYLOG("step");
if (m_genomes.empty())
return;
std::vector<t_genome> bestGenomes;
getBestGenomes(10, bestGenomes);
// D_MYLOG("current_best=" << bestGenomes[0].m_fitness);
// D_MYLOG("step");
// for (t_genome& g : bestGenomes)
// D_MYLOG("g=" << g.m_id);
std::vector<t_genome> children;
children.reserve( m_genomes.size() );
// m_best_fitness = std::max(m_best_fitness, bestGenomes[0].m_fitness);
if (m_best_fitness < bestGenomes[0].m_fitness)
{
m_best_fitness = bestGenomes[0].m_fitness;
D_MYLOG("m_current_generation=" << m_current_generation
<< std::fixed << ", m_best_fitness=" << m_best_fitness);
m_is_a_great_generation = true;
}
else
{
m_is_a_great_generation = false;
}
unsigned int current_index = 0;
if (m_best_fitness > m_alpha_genome.m_fitness)
{
m_alpha_genome = bestGenomes[0];
// D_MYLOG("new alpha");
}
else
{
t_genome bestDude;
bestDude.m_fitness = 0.0f;
bestDude.m_id = m_alpha_genome.m_id;
bestDude.m_weights = m_alpha_genome.m_weights;
bestDude.m_index = current_index++;
mutate(bestDude);
children.push_back(bestDude);
// D_MYLOG("alpha reused");
}
{
// Carry on the best dude.
t_genome bestDude;
bestDude.m_fitness = 0.0f;
bestDude.m_id = bestGenomes[0].m_id;
bestDude.m_weights = bestGenomes[0].m_weights;
bestDude.m_index = current_index++;
mutate(bestDude);
children.push_back(bestDude);
}
// D_MYLOG("step");
// breed best genomes
for (unsigned int i = 0; i < bestGenomes.size(); ++i)
for (unsigned int j = i+1; j < bestGenomes.size(); ++j)
{
// D_MYLOG("i=" << i << ", j=" << j);
t_genome baby1, baby2;
CrossBreed(bestGenomes[i], bestGenomes[j], baby1, baby2);
mutate(baby1);
mutate(baby2);
baby1.m_index = current_index++;
baby2.m_index = current_index++;
children.push_back(baby1);
children.push_back(baby2);
}
// D_MYLOG("step");
// For the remainding n population, add some random kiddies.
unsigned int remainingChildren = (m_genomes.size() - children.size());
// D_MYLOG("step m_genomes.size()" << m_genomes.size());
// D_MYLOG("step children.size()" << children.size());
// D_MYLOG("step remainingChildren" << remainingChildren);
for (unsigned int i = 0; i < remainingChildren; i++)
{
t_genome genome;
genome.m_id = m_current_id++;
genome.m_fitness = 0.0f;
genome.m_weights.resize( m_pNNTopology->getTotalWeights() );
for (float& weight : genome.m_weights)
weight = randomClamped();
genome.m_index = current_index++;
children.push_back( genome );
}
// D_MYLOG("step");
m_genomes = children;
++m_current_generation;
// D_MYLOG("/step");
for (unsigned int i = 0; i < m_genomes.size(); ++i)
m_NNetworks[i].setWeights( m_genomes[i].m_weights );
}
void GeneticAlgorithm::generateRandomPopulation()
{
std::vector<unsigned int> tmp_hidden;
// tmp_hidden.push_back(8);
tmp_hidden.push_back(4);
tmp_hidden.push_back(3);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(6);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(7);
// tmp_hidden.push_back(6);
// tmp_hidden.push_back(5);
// tmp_hidden.push_back(4);
// tmp_hidden.push_back(4);
// tmp_hidden.push_back(3);
m_pNNTopology = t_pNNTopology(new NeuralNetworkTopology(5, tmp_hidden, 2));
// reset the genomes
m_genomes.resize(100);
for (unsigned int i = 0; i < m_genomes.size(); ++i)
{
t_genome& genome = m_genomes[i];
genome.m_index = i;
genome.m_id = m_current_id++;
genome.m_fitness = 0.0f;
genome.m_weights.resize( m_pNNTopology->getTotalWeights() );
for (float& weight : genome.m_weights)
weight = randomClamped();
}
m_NNetworks.reserve( m_genomes.size() );
for (unsigned int i = 0; i < m_genomes.size(); ++i)
{
m_NNetworks.push_back( NeuralNetwork(*m_pNNTopology) );
m_NNetworks[i].setWeights( m_genomes[i].m_weights );
}
m_current_generation = 1;
m_best_fitness = 0.0f;
m_alpha_genome.m_index = -1;
m_alpha_genome.m_id = -1;
m_alpha_genome.m_fitness = 0.0f;
}
