//Rollout function (plays random moves till it hits a termination)
protected override void rollout(AIState rolloutStart)
{
//If the rollout start is a terminal state
int rolloutStartResult = rolloutStart.getWinner();
if (rolloutStartResult >= 0)
{
//Add a win is it is a win, or a loss is a loss or otherwise a draw
if (rolloutStartResult == rolloutStart.playerIndex)
{
rolloutStart.addWin();
}
else if (rolloutStartResult == (rolloutStart.playerIndex + 1) % 2)
{
rolloutStart.addLoss();
}
else
{
rolloutStart.addDraw(drawScore);
}
return;
}
bool terminalStateFound = false;
//Get the children
List <AIState> children = rolloutStart.generateChildren();
int loopCount = 0;
while (!terminalStateFound)
{
//Loop through till a terminal state is found
loopCount++;
//If max roll out is hit or no childern were generated
if (loopCount >= maxRollout || children.Count == 0)
{
//Record a draw
rolloutStart.addDraw(drawScore);
break;
}
//Get a random child index
int index = randGen.Next(children.Count);
//and see if that node is terminal
int endResult = children[index].getWinner();
if (endResult >= 0)
{
terminalStateFound = true;
if (endResult == 2)
{
rolloutStart.addDraw(drawScore);
}
//If it is a win add a win
else if (endResult == rolloutStart.playerIndex)
{
rolloutStart.addWin();
}
//Else add a loss
else
{
rolloutStart.addLoss();
}
}
else
{
//Otherwise select that nodes as the childern and continue
children = children [index].generateChildren();
}
}
//Reset the children as these are not 'real' children but just ones for the roll out.
foreach (AIState child in rolloutStart.children)
{
child.children = new List <AIState>();
}
}
//Rollout function (plays random moves till it hits a termination)
protected override void rollout(AIState rolloutStart)
{
//If the rollout start is a terminal state
int rolloutStartResult = rolloutStart.getWinner();
if (rolloutStartResult >= 0)
{
//Add a win is it is a win, or a loss is a loss or otherwise a draw
if (rolloutStartResult == rolloutStart.playerIndex)
{
rolloutStart.addWin();
}
else if (rolloutStartResult == (rolloutStart.playerIndex + 1) % 2)
{
rolloutStart.addLoss();
}
else
{
rolloutStart.addDraw(drawScore);
}
return;
}
bool terminalStateFound = false;
//Get the children
List <AIState> children = rolloutStart.generateChildren();
int loopCount = 0;
while (!terminalStateFound)
{
//Loop through till a terminal state is found
loopCount++;
//If max roll out is hit or no childern were generated
if (loopCount >= maxRollout || children.Count == 0)
{
//record a draw
rolloutStart.addDraw(drawScore);
break;
}
//Default is the end of the array (because that will be the best move in a sorted list)
int selectedChild = children.Count - 1;
//epsilon greedy move selection.
if (randGen.NextDouble() < epsilon)
{
//Sort the array (we have all ready selected the most move indx above
foreach (AIState child in children)
{
if (child.stateScore == null)
{
child.stateScore = model.evaluate(child);
}
}
children = AIState.mergeSort(children);
}
else
{
//Just select a random move
selectedChild = randGen.Next(children.Count);
}
//and see if that node is terminal
int endResult = children[selectedChild].getWinner();
if (endResult >= 0)
{
terminalStateFound = true;
if (endResult == 2)
{
rolloutStart.addDraw(drawScore);
}
//If it is a win add a win
else if (endResult == rolloutStart.playerIndex)
{
rolloutStart.addWin();
}
//Else add a loss
else
{
rolloutStart.addLoss();
}
}
else
{
//Otherwise select that nodes as the childern and continue
children = children [selectedChild].generateChildren();
}
}
//Reset the children as these are not 'real' children but just ones for the roll out.
foreach (AIState child in rolloutStart.children)
{
child.treeNode = true;
}
}