| public static RandomDouble ( double min, double max ) : double | ||
| min | double | |
| max | double | |
| return | double | |
override public int Search()
{
CtwContextTreeUndo undoInstance = new CtwContextTreeUndo(this);
MonteCarloSearchNode searchTree = new MonteCarloSearchNode(MonteCarloSearchNode.DecisionNode);
for (int i = 0; i < this.McSimulations; i++)
{
searchTree.Sample(this, this.Horizon);
this.model_revert(undoInstance);
}
//searchTree.PrintBs();
int bestAction = -1;
double bestMean = double.NegativeInfinity;
foreach (int action in this.Environment.ValidActions)
{
if (!searchTree.Children.ContainsKey(action))
{
continue;
}
double mean = searchTree.Children[action].Mean + Utils.RandomDouble(0, 0.0001);
if (mean > bestMean)
{
bestMean = mean;
bestAction = action;
}
}
return(bestAction);
}
public int SelectAction(Agent agent)
{
Debug.Assert(agent.MaximumReward() != null, "this is weird place, - in selection action");
double exploreBias = (double)agent.Horizon * agent.MaximumReward().Value;
double explorationNumerator = this.ExplorationConstant * Math.Log(this.Visits);
int bestAction = -1;
double bestPriority = double.NegativeInfinity;
foreach (int action in agent.Environment.ValidActions)
{
MonteCarloSearchNode node = null;
if (this.Children.ContainsKey(action))
{
node = this.Children[action];
}
double priority;
if (node == null || node.Visits == 0)
{
// previously unexplored node
priority = this.UnexploredBias; //unexplored bias
}
else
{
priority = node.Mean + exploreBias * Math.Sqrt(explorationNumerator / node.Visits);
}
if (priority > (bestPriority + Utils.RandomDouble(0, 0.001)))
{
bestAction = action;
bestPriority = priority;
}
}
return(bestAction);
}
override public double Playout(int horizon)
{
return(horizon * Utils.RandomDouble((int)this.Environment.minimum_reward(), (int)this.Environment.maximum_reward()));
}
