PerceptronRef PerceptronGeneticAlgorithmTrainer::Crossover( const PerceptronRef & nw1, const PerceptronRef & nw2 )
{
Perceptron::Layers layers = nw1->GetLayers();
const Perceptron::Layers & layers2 = nw2->GetLayers();
std::uniform_real_distribution<> randDistr;
for ( Perceptron::Layers::size_type layerIndex = 1; layerIndex < layers.size(); ++ layerIndex )
{
Perceptron::Neurons & layer = layers[layerIndex];
const Perceptron::Neurons & layer2 = layers2[layerIndex];
for ( Perceptron::Neurons::size_type neuronIndex = 0; neuronIndex < layer.size(); ++ neuronIndex )
{
Perceptron::Neuron & neuron = layer[neuronIndex];
const Perceptron::Neuron & neuron2 = layer2[neuronIndex];
if ( randDistr( mRandomEngine ) > 0.5 )
{
neuron.threshold = neuron2.threshold;
}
for ( Perceptron::Neuron::Weights::size_type weightIndex = 0; weightIndex < neuron.weights.size(); ++ weightIndex )
{
if ( randDistr( mRandomEngine ) > 0.5 )
{
neuron.weights[weightIndex] = neuron2.weights[weightIndex];
}
}
}
}
return std::make_shared< Perceptron >( layers );
}
PerceptronRef PerceptronGeneticAlgorithmTrainer::Mutation( const PerceptronRef & nw )
{
Perceptron::Layers layers = nw->GetLayers();
std::uniform_real_distribution<> randDistr;
for ( Perceptron::Layers::size_type layerIndex = 1; layerIndex < layers.size(); ++ layerIndex )
{
Perceptron::Neurons & layer = layers[layerIndex];
for ( Perceptron::Neurons::size_type neuronIndex = 0; neuronIndex < layer.size(); ++ neuronIndex )
{
Perceptron::Neuron & neuron = layer[neuronIndex];
if ( randDistr( mRandomEngine ) <= mMutationProbability )
{
neuron.threshold += ( randDistr( mRandomEngine ) - 0.5f ) * mMutationSpeed;
}
for ( Perceptron::Neuron::Weights::size_type weightIndex = 0; weightIndex < neuron.weights.size(); ++ weightIndex )
{
if ( randDistr( mRandomEngine ) <= mMutationProbability )
{
neuron.weights[weightIndex] += ( randDistr( mRandomEngine ) - 0.5f ) * mMutationSpeed;
}
}
}
}
return std::make_shared< Perceptron >( layers );
}
PerceptronRef PerceptronGeneticAlgorithmTrainer::CreateIndividual()
{
Perceptron::Layers layers( 4 );
layers[0].resize( mInputsCount );
layers[1].resize( mInputsCount * 2 );
layers[2].resize( mInputsCount * 2 );
layers[3].resize( mOutputsCount );
std::uniform_real_distribution<> randDistr;
for ( Perceptron::Layers::size_type layerIndex = 1; layerIndex < layers.size(); ++ layerIndex )
{
Perceptron::Neurons & layer = layers[layerIndex];
for ( Perceptron::Neurons::size_type neuronIndex = 0; neuronIndex < layer.size(); ++ neuronIndex )
{
Perceptron::Neuron & neuron = layer[neuronIndex];
neuron.threshold = randDistr( mRandomEngine );
neuron.weights.resize( layers[layerIndex - 1].size() );
for ( Perceptron::Neuron::Weights::size_type weightIndex = 0; weightIndex < neuron.weights.size(); ++ weightIndex)
{
neuron.weights[weightIndex] = randDistr( mRandomEngine );
}
}
}
return std::make_shared< Perceptron >( layers );
}
arma::vec PGPEPolicy::getAction(arma::vec const &observation_) const
{
// Simulate policy parameters
if (randDistr(generator) < resamplingProbability)
policyPtr->setParameters(distributionPtr->simulate());
// Select action
return policyPtr->getAction(observation_);
}
