public override void OnStep(int step, int globalStep, Tensor state, Tensor action, float reward, Tensor nextState, bool done)
{
if (globalStep % MemoryInterval == 0)
{
Memory.Push(new Experience(state, action, reward, nextState, done));
}
if (TargetModelUpdateInterval <= 0)
{
throw new Exception("Target model update has to be positive.");
}
if (TargetModel == null)
{
TargetModel = Net.Clone();
}
if (!TargetModelUpdateOnEpisodeEnd)
{
if (TargetModelUpdateInterval >= 1)
{
if (globalStep % (int)TargetModelUpdateInterval == 0)
{
Net.CopyParametersTo(TargetModel);
}
}
else
{
Net.SoftCopyParametersTo(TargetModel, TargetModelUpdateInterval);
}
}
}
public void NextGeneration()
{
UpdateBestSample();
Net best = best_sample.GetComponentInChildren <NetInterface>().Mind.Clone();
int id = best_sample.GetComponentInChildren <NetInterface>().Id;
float score = best_sample.GetComponentInChildren <NetInterface>().Score;
Destroy(samples[0]);
InstantiateWithNetAndScore(0, best, score, id);
Net net;
for (int i = 1; i < Mathf.FloorToInt(NUMBER_OF_GENERATION_SAMPLES * MUTATION_PERCENTAGE); i++)
{
net = best.Clone();
net.Mutate();
Destroy(samples[i]);
InstantiateWithNet(i, net);
}
for (int i = Mathf.FloorToInt(NUMBER_OF_GENERATION_SAMPLES * MUTATION_PERCENTAGE); i < NUMBER_OF_GENERATION_SAMPLES; i++)
{
net = new Net(
MindConstants.NUMBER_OF_INPUT_NODES,
MindConstants.NUMBER_OF_OUTPUT_NODES,
MindConstants.NUMBER_OF_HIDDEN_NODES,
MindConstants.NUMBER_OF_HIDDEN_LAYERS
);
net.InitializeRandom();
Destroy(samples[i]);
InstantiateWithNet(i, net);
}
}
