public override void TrainModel(List <SARS> batch)
{
var inp = new StatePair[batch.Count];
var outp = new TargetIndexPair[batch.Count];
int i = 0;
foreach (var sars in batch)
{
inp[i] = sars.State.Features;
float target;
if (!sars.NextState.IsTerminal)
{
var a0max = QMax(sars.NextState);
target = sars.Reward + Discount * a0max;
}
else
{
target = sars.Reward;
}
outp[i++] = new TargetIndexPair(target, _amap[sars.Action.ActionId]);
}
for (int j = 0; j < batch.Count; j++)
{
_net.SGD(inp[j], outp[j]);
}
}
public void SGD(StatePair input, TargetIndexPair p)
{
_loss.Clear();
_loss.At(p.Index, p.Target - Compute(input, true)[p.Index]);
var split = _split.Visit(_outback.Visit(_loss, _params), _params);
_backprop.BackPropagation(_unflatten.Visit(split.left, _params), _params);
_hiddenBackprop.Visit(split.right, _params);
IsOutputFromTraining = true;
}
