private void TrainNetwork(Move move)
{
// If there is just one option the training will always reward the NNet
if (move.Children == null)
{
return;
}
if (move.Children.Count < 2)
{
return;
}
if (untrained.Count == 0)
{
UpdateGeneration();
}
JNNet nnet = untrained[random.Next(0, untrained.Count - 1)];
untrained.Remove(nnet);
trained.Add(nnet);
Move nnetMove = null, mostVisitedMove = null;
double score = double.NegativeInfinity, childScore;
int mostVisitedTimes = int.MinValue;
input = State.GetInput();
foreach (Move child in move.Children)
{
childScore = nnet.Next(input)[0];
if (childScore > score)
{
nnetMove = child;
score = childScore;
}
if (child.NumberOfVisits > mostVisitedTimes)
{
mostVisitedMove = child;
mostVisitedTimes = child.NumberOfVisits;
}
}
// Give it a point if it chose the most explored route
if (mostVisitedMove == nnetMove)
{
nnet.Score = 1;
}
else
{
nnet.Score = 0;
}
}
private Move Next()
{
bestNnet = BestNNet;
double score = double.NegativeInfinity, currentScore;
foreach (Move move in State.GetChildren())
{
State.DoMove(move);
currentScore = bestNnet.Next(State.GetInput())[0];
if (currentScore > score)
{
score = currentScore;
bestMove = move;
}
State.UndoMoves(1);
}
return(bestMove);
}
