internal void SaveBestPerformer(int epoch)
{
ITrainingSession bestPerformer = _generation.GetBestPerformer();
var saver = new NeuralNetworkSaver("\\networks");
saver.SaveNeuralNetwork(bestPerformer.NeuralNet, bestPerformer.GetSessionEvaluation(), epoch);
}
public void RunSimulation()
{
RunStarterGeneration();
for (int epoch = 1; epoch < _evolutionConfig.NumEpochs; epoch++)
{
for (int generation = 0; generation < _evolutionConfig.GenerationsPerEpoch; generation++)
{
if (generation == 0)
{
LoggerFactory.GetLogger().Log(LogLevel.Info, $"Creating next generation with top performer with eval: {_bestPerformerOfEpoch.GetSessionEvaluation()}");
createNextGeneration(_bestPerformerOfEpoch);
}
else
{
createNextGeneration(null);
}
_generation.Run();
IEnumerable <double> evals = _generation.GetEvalsForGeneration().OrderByDescending(d => d);
var evalsTotalCount = evals.Count();
if (_evolutionConfig.NumTopEvalsToReport > 0)
{
evals = evals.Take(_evolutionConfig.NumTopEvalsToReport);
}
var sb = new StringBuilder();
foreach (var t in evals.OrderBy(d => d))
{
sb.AppendLine($"eval: {t}");
}
LoggerFactory.GetLogger().Log(LogLevel.Info, sb.ToString());
LoggerFactory.GetLogger().Log(LogLevel.Info, $"count: {evalsTotalCount}");
LoggerFactory.GetLogger().Log(LogLevel.Info, $"Epoch: {epoch}, Generation: {generation}");
}
if (_epochAction != null)
{
LoggerFactory.GetLogger().Log(LogLevel.Info, "Updating best performer");
_bestPerformerOfEpoch = _epochAction.UpdateBestPerformer(_generation, epoch);
}
else
{
_bestPerformerOfEpoch = GetBestPerformerOfGeneration();
}
SaveBestPerformer(epoch);
}
}
private void createNextGeneration(ITrainingSession bestPerformer)
{
Stopwatch watch = new Stopwatch();
watch.Start();
int numberOfTopPerformersToChoose = (int)(_generationConfig.GenerationPopulation * 0.50);
int numToBreed = (int)(_generationConfig.GenerationPopulation * 0.35);
int numToGen = (int)(_generationConfig.GenerationPopulation * 0.15);
if (numberOfTopPerformersToChoose + numToBreed + numToGen < _generationConfig.GenerationPopulation)
{
numToGen += _generationConfig.GenerationPopulation -
(numberOfTopPerformersToChoose + numToBreed + numToGen);
}
var sessions = _generation.GetBestPerformers(numberOfTopPerformersToChoose);
if (bestPerformer != null)
{
LoggerFactory.GetLogger().Log(LogLevel.Info, "Best performer found for creating generation");
if (sessions.All(s => s.NeuralNet.GetGenes() != bestPerformer.NeuralNet.GetGenes()))
{
LoggerFactory.GetLogger()
.Log(LogLevel.Info,
$"Best performer adding to sessions with eval {bestPerformer.GetSessionEvaluation()}");
sessions[sessions.Count - 1] = bestPerformer;
sessions = sessions.OrderByDescending(s => s.GetSessionEvaluation()).ToList();
LoggerFactory.GetLogger()
.Log(LogLevel.Info, $"session 0 eval: {sessions[0].GetSessionEvaluation()}");
}
else
{
LoggerFactory.GetLogger()
.Log(LogLevel.Info,
$"Best performer already in generation: not adding.");
}
}
_history.AddEval(sessions[0].GetSessionEvaluation());
var mutateChance = DetermineMutateChance();
IList <INeuralNetwork> children = _breeder.Breed(sessions, numToBreed);
var newSessions = new List <ITrainingSession>();
//Allow the very top numToLiveOn sessions to be added to next generation untouched
int numToLiveOn = sessions.Count / 10;
var sessionsToLiveOn = sessions.Take(numToLiveOn).ToList();
newSessions.AddRange(sessionsToLiveOn);
//try to mutate session that will live on 1 by 1
var mutatedTop = GetMutatedNetworks(sessionsToLiveOn.Select(s => s.NeuralNet), mutateChance);
//or each session that lived on that was mutated, remove last top performer, then mutate remaining top performers in batch and add
sessions = sessions.Skip(numToLiveOn).ToList();
var sessionSubset = sessions.Take(sessions.Count - mutatedTop.Count);
IList <INeuralNetwork> toKeepButPossiblyMutate = sessionSubset.Select(session => session.NeuralNet).ToList();
IList <INeuralNetwork> newNetworks = getNewNetworks(numToGen, mutateChance);
List <INeuralNetwork> toTryMutate = new List <INeuralNetwork>();
//try to mutate both new networks as well as all the top performers we wanted to keep
toTryMutate.AddRange(toKeepButPossiblyMutate);
toTryMutate.AddRange(newNetworks);
bool didMutate;
IList <INeuralNetwork> maybeMutated = _mutator.Mutate(toTryMutate, mutateChance, out didMutate);
List <INeuralNetwork> allToAdd = new List <INeuralNetwork>();
allToAdd.AddRange(mutatedTop);
allToAdd.AddRange(children);
allToAdd.AddRange(maybeMutated);
newSessions.AddRange(allToAdd.Select((net, sessionNumber) => new TrainingSession(net, _evaluatableFactory.Create(net), sessionNumber)));
_generation = new Generation(newSessions, _generationConfig);
watch.Stop();
LoggerFactory.GetLogger().Log(LogLevel.Debug, $"create generation runtime (sec): {watch.Elapsed.TotalSeconds}");
watch.Reset();
}