// Use this for initialization
void Start()
{
PPONetworkContinuousSimple network;
if (environment.is3D)
{
network = new PPONetworkContinuousSimple(8, 2, 2, 32, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 10000, 200);
trainer.ClipEpsilon = 0.1f;
}
else
{
network = new PPONetworkContinuousSimple(5, 2, 2, 32, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 10000, 200);
}
//test
//trainer.RewardDiscountFactor = 0.5f;
loss = new AutoAverage(iterationForEachTrain);
episodePointAve = new AutoAverage(episodeToRunForEachTrain);
}
// Use this for initialization
void Start()
{
environment.isContinuou = false;
QNetworkSimple network = new QNetworkSimple(2, 2, 2, 64, DeviceDescriptor.CPUDevice, 0.4f);
model = new DQLModel(network);
//QNetworkSimple networkTarget = new QNetworkSimple(2, 2, 2, 64, DeviceDescriptor.CPUDevice, 1f);
//modelTarget = new DQLModel(networkTarget);
//trainer = new TrainerDQLSimple(model, null, LearnerDefs.MomentumSGDLearner(startLearningRate,0.9f),1, experienceBufferSize, 500);
trainer = new TrainerDQLSimple(model, modelTarget, LearnerDefs.AdamLearner(startLearningRate), 1, experienceBufferSize, 500);
//Save();//test
}
// Use this for initialization
void Start()
{
QNetworkSimple network = new QNetworkSimple(6, 3, 2, 64, DeviceDescriptor.GPUDevice(0), 0.4f);
model = new DQLModel(network);
QNetworkSimple networkTarget = new QNetworkSimple(6, 3, 2, 64, DeviceDescriptor.GPUDevice(0), 0.4f);
modelTarget = new DQLModel(networkTarget);
//trainer = new TrainerDQLSimple(model, null, LearnerDefs.SGDLearner(startLearningRate),1, experienceBufferSize, 2048);
trainer = new TrainerDQLSimple(model, modelTarget, LearnerDefs.AdamLearner(startLearningRate), 1, experienceBufferSize, experienceBufferSize);
//Save();//test
}
// Use this for initialization
void Start()
{
var network = new PPONetworkDiscreteSimple(6, 3, 4, 64, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 50000, 2000);
//test
//trainer.RewardDiscountFactor = 0.5f;
loss = new AutoAverage(iterationForEachTrain);
episodePointAve = new AutoAverage(episodeToRunForEachTrain);
}
// Use this for initialization
void Start()
{
//QNetworkSimple network = new QNetworkSimple(environment.mazeDimension.x* environment.mazeDimension.y, 4, 3, 64, DeviceDescriptor.CPUDevice, 0.4f);
var network = new QNetworkConvSimple(environment.mazeDimension.x, environment.mazeDimension.y, 1,
4, new int[] { 3, 3 }, new int[] { 64, 128 }, new int[] { 1, 1 }, new bool[] { true, true }, 1, 128,
false, DeviceDescriptor.GPUDevice(0), 1f);
model = new DQLModel(network);
//QNetworkSimple networkTarget = new QNetworkSimple(environment.mazeDimension.x * environment.mazeDimension.y, 4, 3, 64, DeviceDescriptor.CPUDevice, 0.4f);
//modelTarget = new DQLModel(networkTarget);
//trainer = new TrainerDQLSimple(model, modelTarget, LearnerDefs.SGDLearner(startLearningRate), 1, experienceBufferSize, 500);
trainer = new TrainerDQLSimple(model, null, LearnerDefs.SGDLearner(startLearningRate), 1, experienceBufferSize, 500);
//Save();//test
}
public void CreateResnode()
{
var input = new InputLayerDense(2);
//outputLayer = new OutputLayerDenseBayesian(1);
outputLayer = new OutputLayerDense(1, null, OutputLayerDense.LossFunction.Square);
network = new SequentialNetworkDense(input, LayerDefineHelper.DenseLayers(10, 5, true, NormalizationMethod.None), outputLayer, DeviceDescriptor.CPUDevice);
//network = new SequentialNetworkDense(input, LayerDefineHelper.ResNodeLayers(10, 5), outputLayer, DeviceDescriptor.CPUDevice);
trainer = new TrainerSimpleNN(network, LearnerDefs.AdamLearner(lr), DeviceDescriptor.CPUDevice);
dataPlane.network = this;
}
public void CreateGAN()
{
gan = new GAN(2, 0, 2, 10, 5, 10, 5, DeviceDescriptor.GPUDevice(0));
trainerGan = new TrainerGAN(gan, LearnerDefs.AdamLearner(lrGenerator), LearnerDefs.AdamLearner(lrDiscriminator), DeviceDescriptor.GPUDevice(0));
trainerGan.usePredictionInTraining = true;
}
