/// <summary>
/// OOP advatages adopted during translation...
/// </summary>
/// <param name="experience">See Experience</param>
/// <returns></returns>
private double LearnFromExperience(Experience experience /*Matrix s0, int a0, double r0, Matrix s1, int a1*/)
{
// want: Q(s,a) = r + gamma * max_a' Q(s',a')
// compute the target Q value
var tmat = ForwardQ(Network, s1, false);
var qmax = r0 + Options.Gamma * tmat.W[Tembo.Maxi(tmat.W)];
// now predict
var pred = ForwardQ(Network, s0, true);
var tderror = pred.W[a0] - qmax;
var clamp = Options.ErrorClamp;
if (Math.Abs(tderror) > clamp)
{ // huber loss to robustify
if (tderror > clamp)
{
tderror = clamp;
}
if (tderror < -clamp)
{
tderror = -clamp;
}
}
pred.DW[a0] = tderror;
LastGraph.Backward(); // compute gradients on net params
// update net
Tembo.UpdateNetwork(Network, Options.Alpha);
return(tderror);
}
