private MCTSAZNode Select(MCTSAZNode root)
{
while (root.Children.Count != 0) //Look for a leaf node (one we haven't expanded yet)
{
MCTSAZNode bestNext = null;
double bestValue = double.MinValue;
for (var i = 0; i < root.Children.Count; i++)
{
var c = root.Children[i];
double uctValue = c.AverageValue + c.CalculateUCT(_random, root, root.NetworkResult.MoveRating[c.NetworkChildIndex]);
// small random number to break ties randomly in unexpanded nodes
if (root.Parent == null)
{
uctValue *= _random.NextDouble();
}
if (uctValue > bestValue)
{
bestNext = c;
bestValue = uctValue;
}
}
root = bestNext;
}
return(root);
}
public double CalculateUCT(Random random, MCTSAZNode parent, float priorProbability)
{
//https://web.stanford.edu/~surag/posts/alphazero.html
//Page 8 https://www.nature.com/articles/nature24270.epdf?author_access_token=VJXbVjaSHxFoctQQ4p2k4tRgN0jAjWel9jnR3ZoTv0PVW4gB86EEpGqTRDtpIz-2rmo8-KG06gqVobU5NSCFeHILHcVFUeMsbvwS-lxjqQGg98faovwjxeTUgZAUMnRQ
return(ExplorationParameter * (priorProbability + Epsilon) * Math.Sqrt(parent.VisitCount + 1) / (1 + VisitCount));
}
public void Reset()
{
Children.Clear();
Value = 0;
VisitCount = 0;
AverageValue = 0;
Parent = null;
NetworkResult = null;
PieceToPurchase = null;
NetworkChildIndex = -1;
}
/// <summary>
/// Finds the best child based on their VisitCount
/// </summary>
/// <returns></returns>
internal MCTSAZNode FindBestChild(MCTSAZNode root)
{
//Perform the best move
var best = root.Children[0];
int bestVisitCount = root.Children[0].VisitCount;
for (var index = 1; index < root.Children.Count; index++)
{
var child = root.Children[index];
if (child.VisitCount > bestVisitCount) //TODO: Handle draws
{
best = child;
bestVisitCount = child.VisitCount;
}
}
return(best);
}
private void Expand(MCTSAZNode root)
{
root.Expand(_client);
}
