public override HashSet <A> PerformActions(Dictionary <string, float> inputs)
{
Dictionary <string, float> outputs = SequentialNN.Predict(inputs);
KeyValuePair <string, float>?maxOuput = null;
foreach (KeyValuePair <string, float> ouput in outputs)
{
if (maxOuput == null)
{
maxOuput = ouput;
}
else if (ouput.Value > maxOuput.Value.Value)
{
maxOuput = ouput;
}
}
A action = (A)Enum.Parse(typeof(A), maxOuput.Value.Key);
HashSet <A> actions = new HashSet <A>();
actions.Add(action);
return(actions);
}
public override EvolutionalEntity <A> CreateOffspring()
{
SequentialNN childSequentialNN = new SequentialNN(SequentialNN);
//mutate child
foreach (Layer layer in childSequentialNN.Layers)
{
foreach (Perceptron perceptron in layer.Perceptrons)
{
float randomModifier;
foreach (KeyValuePair <string, float> weight in perceptron.Weights.ToList())
{
randomModifier = (float)(random.NextDouble() * 2 - 1) * LearningTolerance;
perceptron.Weights[weight.Key] = weight.Value + randomModifier;
}
randomModifier = (float)(random.NextDouble() * 2 - 1) * LearningTolerance;
perceptron.BiasWeight += randomModifier;
}
}
//float learningTolerance = LearningTolerance - 0.01f;
//if (learningTolerance < 0)
//{
// learningTolerance = 0;
//}
//return new NNEvolutionalEntity<A>(childSequentialNN) { LearningTolerance = learningTolerance };
return(new NNEvolutionalEntity <A>(childSequentialNN));
}
public StockNNES(SequentialNN model, int populationSize)
{
State defaultState = new State();
defaultState.Cash = 1000;
defaultState.Stocks = 0;
DefaultState = defaultState;
CreateInitialPopulation(model, populationSize);
}
private void CreateInitialPopulation(SequentialNN model, int populationSize)
{
//create initial population
Population = new List <EvolutionalEntity <Action> >();
for (int i = 0; i < populationSize; i++)
{
//copy population based on model
SequentialNN newNN = new SequentialNN();
foreach (Layer layer in model.Layers)
{
Layer newLayer = new Layer(layer.Perceptrons.Count);
newNN.AddLayer(newLayer);
}
newNN.AddLayer(new Layer(typeof(Action).GetEnumNames().Count()));
NNEvolutionalEntity <Action> entity = new NNEvolutionalEntity <Action>(newNN);
Population.Add(entity);
States.Add(entity, DefaultState);
}
}
static void Main(string[] args)
{
const int TICKS_TRAIN = 120;
const int TICKS_TEST = 120;
const int INPUT_HOURS = 24;
const int GENERATION_TO_STOP_TRAINING = 50;
const int POPULATION_SIZE = 100;
StockEnvironment stockEnvironment;
//build model
SequentialNN model = new SequentialNN();
model.AddLayer(new Layer(5));
model.AddLayer(new Layer(5));
StockNNES stockNNES = new StockNNES(model, POPULATION_SIZE);
StockNNES stockNNESRandom = new StockNNES(model, POPULATION_SIZE);
//train
for (int i = 0; i < GENERATION_TO_STOP_TRAINING; i++)
{
Console.Clear();
Console.WriteLine("Training Generation: {0}", stockNNES.Generation);
//create new environment
stockEnvironment = new StockEnvironment("XVG", INPUT_HOURS, TICKS_TRAIN, TICKS_TEST + INPUT_HOURS - 1);
stockNNES.Environment = stockEnvironment;
for (int j = 0; j < TICKS_TRAIN; j++)
{
stockNNES.RunTick();
}
if (i != GENERATION_TO_STOP_TRAINING - 1)
{
stockNNES.CreateNextGeneration();
}
}
//print performances
Console.WriteLine();
Console.WriteLine("Performance of trained population");
Console.WriteLine("----------------");
PrintPerformance(stockNNES);
EvolutionalEntity <Action> bestTrainingPerformer = stockNNES.Performances.OrderBy(p => p.Value).Last().Key;
//reset performance
stockNNES.ResetState();
//test trained
Console.WriteLine("Testing Trained");
stockEnvironment = new StockEnvironment("BTC", INPUT_HOURS, TICKS_TEST);
stockNNES.Environment = stockEnvironment;
for (int i = 0; i < TICKS_TEST; i++)
{
stockNNES.RunTick();
}
//test random
Console.WriteLine("Testing Random");
stockEnvironment = new StockEnvironment("BTC", INPUT_HOURS, TICKS_TEST);
stockNNESRandom.Environment = stockEnvironment;
for (int i = 0; i < TICKS_TEST; i++)
{
stockNNESRandom.RunTick();
}
//print performances of random population
Console.WriteLine();
Console.WriteLine("Performance of random population");
Console.WriteLine("----------------");
PrintPerformance(stockNNESRandom);
//print performances
Console.WriteLine();
Console.WriteLine("Performance of trained population");
Console.WriteLine("----------------");
PrintPerformance(stockNNES, bestTrainingPerformer);
}
public NNEvolutionalEntity(SequentialNN sequentialNN)
{
SequentialNN = sequentialNN;
LearningTolerance = 1;
}
public TrainReturn Train(TrainParameters trainParameters)
{
TrainReturn trainReturn = new TrainReturn();
List <NNStockStatistics> trainStatistics = new List <NNStockStatistics>();
List <NNStockStatistics> testStatistics = new List <NNStockStatistics>();
List <NNStockStatistics> trainAverage = new List <NNStockStatistics>();
List <NNStockStatistics> testAverage = new List <NNStockStatistics>();
SequentialNN model = new SequentialNN();
model.AddLayer(new Layer(5));
model.AddLayer(new Layer(5));
StockNNES stockNNES = new StockNNES(model, 100);
for (int i = 0; i < trainParameters.generations; i++)
{
//train
stockNNES.Environment = new StockEnvironment(trainParameters.stockName, trainParameters.inputSize, trainParameters.ticksTrain, trainParameters.ticksTest + trainParameters.inputSize - 1);
for (int j = 0; j < trainParameters.ticksTrain; j++)
{
stockNNES.RunTick();
}
KeyValuePair <EvolutionalEntity <Action>, float> bestPerformerTrain = stockNNES.Performances.OrderBy(p => p.Value).Last();
trainStatistics.Add(new NNStockStatistics {
EvolutionalEntity = bestPerformerTrain.Key, Performance = bestPerformerTrain.Value, Generation = i + 1
});
trainAverage.Add(new NNStockStatistics {
EvolutionalEntity = null, Performance = GetAverage(stockNNES), Generation = i + 1
});
Dictionary <EvolutionalEntity <Action>, float> performance = new Dictionary <EvolutionalEntity <Action>, float>(stockNNES.Performances);
//test
stockNNES.Environment = new StockEnvironment(trainParameters.stockName, trainParameters.inputSize, trainParameters.ticksTest);
for (int j = 0; j < trainParameters.ticksTest; j++)
{
stockNNES.RunTick();
}
KeyValuePair <EvolutionalEntity <Action>, float> bestPerformerTest = stockNNES.Performances.OrderBy(p => p.Value).Last();
testStatistics.Add(new NNStockStatistics {
EvolutionalEntity = bestPerformerTest.Key, Performance = bestPerformerTest.Value, Generation = i + 1
});
testAverage.Add(new NNStockStatistics {
EvolutionalEntity = null, Performance = GetAverage(stockNNES), Generation = i + 1
});
GetAverage(stockNNES);
//reset performance
stockNNES.Performances = performance;
if (i != trainParameters.generations - 1)
{
stockNNES.CreateNextGeneration();
}
}
trainReturn.trainStatistics = trainStatistics;
trainReturn.testStatistics = testStatistics;
trainReturn.trainAverage = trainAverage;
trainReturn.testAverage = testAverage;
return(trainReturn);
}
