public GOB.Action Run()
{
//Default actions that should always be taken
Action a = CheckAlwaysBestAction(InitialNodes[0]);
if (a != null)
{
return(a);
}
MCTSNode selectedNode;
Reward reward;
var startTime = Time.realtimeSinceStartup;
this.CurrentIterationsInFrame = 0;
int currentMCTS = 0;
while (CurrentIterations < MaxIterations)
{
if (CurrentIterationsInFrame >= MaxIterationsProcessedPerFrame)
{
TotalProcessingTime += Time.realtimeSinceStartup - startTime;
return(null);
}
selectedNode = Selection(InitialNodes[currentMCTS]);
if (selectedNode == InitialNodes[currentMCTS])
{
break;
}
reward = Playout(selectedNode.State);
Backpropagate(selectedNode, reward);
CurrentIterationsInFrame++;
CurrentIterations++;
currentMCTS++;
if (currentMCTS == NumberOfRuns)
{
currentMCTS = 0;
}
}
BestFirstChild = BestChildFromSeveral(InitialNodes);
MCTSNode child = BestFirstChild;
BestActionSequence.Clear();
while (child != null)
{
BestActionSequence.Add(child.Action);
child = BestChild(child);
}
InProgress = false;
if (BestFirstChild != null)
{
TotalProcessingTime += Time.realtimeSinceStartup - startTime;
return(BestFirstChild.Action);
}
return(null);
}
public GOB.Action Run()
{
//TODO: Execute MCTS PlayoutIterations times
var startTime = Time.realtimeSinceStartup;
this.CurrentIterationsInFrame = 0;
int MaxIterations = this.MaxIterationsProcessedPerFrame / this.PlayoutIterations;
MCTSNode selectedNode = new MCTSNode(this.CurrentStateWorldModel.GenerateChildWorldModel());
List <float> results = new List <float>();
int i = 0;
for (i = 0; i < this.PlayoutIterations; i++)
{
Reward reward;
while (this.CurrentIterationsInFrame < MaxIterations)
{
MCTSNode newNode = Selection(selectedNode);
reward = Playout(newNode.State);
Backpropagate(newNode, reward);
this.CurrentIterationsInFrame++;
}
for (int j = 0; j < selectedNode.ChildNodes.Count; j++)
{
if (results.Count <= j)
{
results.Add(selectedNode.ChildNodes[j].Q / selectedNode.ChildNodes[j].N);
}
else
{
results[j] += selectedNode.ChildNodes[j].Q / selectedNode.ChildNodes[j].N;
}
if (i == this.PlayoutIterations - 1)
{
results[j] /= this.PlayoutIterations;
}
}
}
this.TotalProcessingTime += Time.realtimeSinceStartup - startTime;
this.InProgress = false;
//MCTSNode best = BestChild(selectedNode);
MCTSNode best = BestAverageChild(selectedNode, results);
BestActionSequence.Clear();
GOB.Action bestAction = best != null ? best.Action : null;
if (bestAction != null)
{
BestActionSequence.Add(bestAction);
}
return(bestAction);
}
//this method is very similar to the bestUCTChild, but it is used to return the final action of the MCTS search, and so we do not care about
//the exploration factor
private MCTSNode BestChild(MCTSNode node)
{
if (node.ChildNodes.Count == 0)
{
return(null);
}
float A = C * (float)Math.Sqrt(node.N);
MCTSNode bestChild = null;
float bestReward = 0;
for (int i = 0; i < node.ChildNodes.Count; i++)
{
float newReward = 0;
switch (strategy)
{
case BestStrategy.Max:
newReward = node.ChildNodes[i].Q / node.ChildNodes[i].N;
break;
case BestStrategy.Robust:
newReward = node.ChildNodes[i].N;
break;
case BestStrategy.MaxRobust:
newReward = node.ChildNodes[i].Q / node.ChildNodes[i].N + node.ChildNodes[i].N;
break;
case BestStrategy.Secure:
newReward = node.ChildNodes[i].Q / node.ChildNodes[i].N - A / Mathf.Sqrt(node.ChildNodes[i].N);
break;
}
if (newReward > bestReward || bestChild == null)
{
bestChild = node.ChildNodes[i];
bestReward = newReward;
}
}
if (bestChild == null || bestChild.Action == null)
{
BestActionSequence.Clear();
}
return(bestChild);
}
private MCTSNode BestAverageChild(MCTSNode node, List <float> results)
{
MCTSNode bestChild = null;
float bestReward = 0;
for (int i = 0; i < results.Count; i++)
{
float newReward = results[i];
if (newReward > bestReward || bestChild == null)
{
bestChild = node.ChildNodes[i];
bestReward = newReward;
}
}
if (bestChild == null || bestChild.Action == null)
{
BestActionSequence.Clear();
}
return(bestChild);
}
public GOB.Action Run()
{
Reward reward;
this.CurrentIterationsInFrame = 0;
MCTSNode initialNode = this.InitialNode, selectedNode;
while (CurrentIterations < MaxIterations && CurrentIterationsInFrame < MaxIterationsProcessedPerFrame)
{
selectedNode = Selection(initialNode);
reward = Playout(selectedNode.State);
Backpropagate(selectedNode, reward);
CurrentIterations++;
CurrentIterationsInFrame++;
}
if (CurrentIterations >= MaxIterations)
{
InProgress = false;
}
TotalProcessingTime += Time.deltaTime;
BestActionSequence.Clear();
BestFirstChild = BestChild(initialNode);
if (BestFirstChild == null)
{
return(null);
}
MCTSNode child = BestFirstChild;
while (child != null)
{
BestActionSequence.Add(child.Action);
child = BestChild(child);
}
return(BestFirstChild.Action);
}