public NeuralAgentTrainer(NeuralAgentTrainerConfiguration configuration)
{
NetworkName = $"Auto-R{configuration.ViewRadius}-{string.Join("_", configuration.LayersSizes)}";
InitialiseNetwork(NetworkName, configuration);
TrainingPluginConfiguration = configuration.TrainingPluginConfiguration;
TrainingSimulationConfiguration = configuration.TrainingSimulationConfiguration;
TrainingProgressTrackerConfiguration = configuration.TrainingProgressTrackerConfiguration;
TestingSimulationConfiguration = configuration.TrainingProgressSimulationConfiguration;
SpaceTemplateFactory = Workspace.Instance.Spaces.GetByName(configuration.Space);
TrainingOuputDirectory = Workspace.Instance.NeuralTrainingsDirectory.Ensure().CreateSubdirectory($"{NetworkName}-{DateTime.Now:yyyy_MM_dd-HH_mm}");
WriteLine("Training Plugin Configuration");
WriteLine("-----------------------------");
WriteLine(TrainingPluginConfiguration.ConvertToJson());
WriteLine("");
WriteLine("Training Simulation Configuration");
WriteLine("---------------------------------");
WriteLine(TrainingSimulationConfiguration.ConvertToJson());
WriteLine("");
WriteLine("Training Progress Tracker Configuration");
WriteLine("---------------------------------------");
WriteLine(TrainingProgressTrackerConfiguration.ConvertToJson());
WriteLine("");
WriteLine("Training Simulation Configuration");
WriteLine("---------------------------------");
WriteLine(TestingSimulationConfiguration.ConvertToJson());
WriteLine("");
}
public NeuralSimulationPlugin(FileInfo networkFile, AgentNetworkTrainingConfiguration trainingConfiguration, int trainingSeed) : base(new NeuralAgentFactory())
{
TrainingRandom = new Random(trainingSeed);
NetworkFile = networkFile;
Network = new AgentNetwork(networkFile);
TrainingConfiguration = trainingConfiguration;
}
public void Fit(IEnumerable <MarkovHistoryItem> batch, AgentNetworkTrainingConfiguration configuration, Random random)
{
var optimizer = new SGDMomentum(Network, configuration.LearningRate, configuration.Momentum);
var trainer = new Trainer(optimizer, random);
var nextQ = batch.Select(item => new Projection(item.Input, new double[] { item.Reward + configuration.Gamma * Predict(item.State).Value })).ToArray();
trainer.Train(nextQ, configuration.EpochesPerIteration, configuration.BatchSize);
}
public NeuralSimulationPluginFactory(string name, string network, int trainingSeed)
{
Name = name;
Network = network;
TrainingSeed = trainingSeed;
TrainingConfiguration = new AgentNetworkTrainingConfiguration();
TrainingProgressTrackerConfiguration = new TrainingProgressTrackerConfiguration();
}
public NeuralAgentTrainerConfiguration()
{
TrainingPluginConfiguration = new AgentNetworkTrainingConfiguration();
TrainingSimulationConfiguration = new SimulationModelConfiguration();
TrainingProgressTrackerConfiguration = new TrainingProgressTrackerConfiguration();
TrainingProgressSimulationConfiguration = new SimulationModelConfiguration();
TrainingSimulationConfiguration.AgentDestructionModel.RemoveDestoryed = true;
TrainingSimulationConfiguration.IterationLimit = 50000;
TrainingProgressSimulationConfiguration.AgentDestructionModel.RemoveDestoryed = true;
}
