private MctsNode createNode(BoardState state, MctsNode parent, int incomingMove)
{
GameObject gob = Instantiate(m_nodePrefab.gameObject);
MctsNode node = gob.GetComponent <MctsNode>();
node.m_unexpandedMoves = new List <int>();
for (int i = 0; i < 9; i++)
{
if (state.m_cellState[i] == 0)
{
node.m_unexpandedMoves.Add(i);
}
}
node.m_incomingMove = incomingMove;
node.m_incomingMovePlayer = 3 - state.m_currentPlayer;
node.m_parent = parent;
if (parent != null)
{
parent.m_children.Add(node);
}
if (parent != null)
{
node.m_depth = parent.m_depth + 1;
}
else
{
node.m_depth = 0;
}
Vector3 position;
if (parent == null)
{
position = transform.position + Vector3.back * transform.localScale.z;
}
else
{
position = parent.transform.position + Vector3.forward * transform.localScale.z;
int cx = (incomingMove % 3) - 1;
int cy = (incomingMove / 3) - 1;
position.x += cx * transform.localScale.x * Mathf.Pow(0.25f, node.m_depth - 1);
position.y += cy * transform.localScale.y * Mathf.Pow(0.25f, node.m_depth - 1);
//const float expansion = 1.1f;
//position.x *= Mathf.Pow(expansion, node.m_depth - 1);
//position.y *= Mathf.Pow(expansion, node.m_depth - 1);
}
node.setPosition(position);
m_nodes.Add(node);
return(node);
}
private IEnumerator doMCTS()
{
clearNodes();
MctsNode rootNode = createNode(m_state, null, -1);
for (int t = 0; t < m_iterations; t++)
{
MctsNode currentNode = rootNode;
BoardState currentState = new BoardState(m_state);
// Selection
while (!currentState.IsTerminal && currentNode.m_unexpandedMoves.Count == 0)
{
currentNode = currentNode.selectChild();
currentState.playMove(currentNode.m_incomingMove);
}
// Expansion
if (!currentState.IsTerminal)
{
int index = Random.Range(0, currentNode.m_unexpandedMoves.Count);
int move = currentNode.m_unexpandedMoves[index];
currentNode.m_unexpandedMoves.RemoveAt(index);
currentState.playMove(move);
currentNode = createNode(currentState, currentNode, move);
}
// Simulation
while (!currentState.IsTerminal)
{
int index = Random.Range(0, 9);
currentState.playMove(index);
}
// Backpropagation
for (; currentNode != null; currentNode = currentNode.m_parent)
{
currentNode.updateStatistics(currentState.m_terminalReward);
}
if (t % m_iterationsPerFrame == 0)
{
yield return(null);
}
}
int chosenMove = rootNode.getMostVisitedChild().m_incomingMove;
m_state.playMove(chosenMove);
updateBoard();
m_waitingForPlayerMove = true;
}
public MctsNode selectChild()
{
float bestScore = 0;
MctsNode bestChild = null;
foreach (MctsNode child in m_children)
{
float score = child.calculateUcb() + Random.Range(0, 1.0e-5f);
if (bestChild == null || score > bestScore)
{
bestChild = child;
bestScore = score;
}
}
return(bestChild);
}
public MctsNode getMostVisitedChild()
{
float bestScore = 0;
MctsNode bestChild = null;
foreach (MctsNode child in m_children)
{
float score = child.m_visits + Random.Range(0, 1.0e-5f);
if (bestChild == null || score > bestScore)
{
bestChild = child;
bestScore = score;
}
}
return(bestChild);
}