public override QState Run(QState currentState, int trialNum, decimal learn, decimal discount, decimal explore)
{
QSearch qsearch = new QSearch(this);
QSearchResult actions = qsearch.Breadth_First(currentState, true);
if (actions != null)
{
foreach (QAction action in actions.actionsList)
{
if (!currentState.IsEnd() && isRunning && currentState.GetActions().Contains(action))
{
WriteOutput(currentState + ": " + action);
QState newState = currentState.GetNewState(action);
newState.Inherit(currentState);
newState.Step();
currentState = newState;
}
}
if (currentState.IsEnd()) WriteOutput(currentState + ": End");
else
{
WriteOutput("Existing solution no longer applicable. Re-solving...");
return Run(currentState, trialNum, learn, discount, explore);
}
}
else WriteOutput("No solution found.", true);
return currentState;
}
public override QState Initialize()
{
WriteOutput("Dimensions: " + width + "x" + height);
self = new Point(start.X, start.Y);
score = 0;
walls = new Point[] { };
// Generate Walls Randomly
if (wallDensity > 0)
{
HashSet<Point> bestPathSoFar = new HashSet<Point>();
WriteOutput("Generating walls randomly with density target of " + wallDensity + "...");
for (int w = 0; w < width; w++)
{
for (int h = 0; h < height; h++)
{
double r = random.NextDouble();
//WriteOutput("Wall Probability for " + w + "x" + h + ": " + r+" vs threshold "+walls);
if (r < wallDensity)
{
//WriteOutput("Wall created at " + w + "x" + h);
Point newWall = new Point(w, h);
if (start == newWall || goal == newWall) continue;
Point[] tempWalls = walls.Concat(new Point[] { newWall }).ToArray();
QState tempState = new Maze() { maze = maze, self = self, goal = goal, width = width, height = height, walls = tempWalls};
if (!bestPathSoFar.Any() || bestPathSoFar.Contains(newWall))
{
QSearchResult path = new QSearch().AStar(tempState);
if (path != null)
{
bestPathSoFar.Clear();
foreach (QState q in path.QStatesList)
bestPathSoFar.Add(((Maze)q).self);
walls = tempWalls;
}
}
else walls = tempWalls;
}
}
}
WriteOutput("Maze generation complete.");
}
opponent = new List<Point>();
for (int i = 0; i < opponents; i++)
opponent.Add(new Point(goal.X, goal.Y));
if (!HideOutput)
{
ManualResetEvent wait = new ManualResetEvent(false);
ThreadPool.QueueUserWorkItem(new WaitCallback(CreateGUI), new object[] { width, height, self.X, self.Y, goal.X, goal.Y, walls, this, wait });
wait.WaitOne();
}
return new Maze() { width = width, height = height, self = self, goal = goal, walls = walls, wallDensity = wallDensity, opponent = opponent, opponentDifficulty = opponentDifficulty, random = random, maze = maze, start = start, opponents = opponents, bestOpponentPath = bestOpponentPath, score = score };
}
public override Dictionary<QFeature, decimal> GetFeatures(QAction action)
{
Point self = ((Maze_With_No_Walls)GetNewState(action)).self;
QSearch qsearch = new QSearch(this);
Maze_With_No_Walls simpleMaze = new Maze_With_No_Walls() { maze = maze, self = self, goal = goal, width = width, height = height };
QSearchResult bestPath = qsearch.AStar(simpleMaze);
return QFeature_String.FromStringDecimalDictionary(new Dictionary<string, decimal>() {
//{this.GetHashCode().ToString()+"_"+action, 1}, // Identity to convert this back to QLearning
{"Goal", goal==self? 1:0},
{"Distance_To_Goal", bestPath==null? 1:(decimal) bestPath.Count / (decimal)(width * height)},
});
}
public override Dictionary<QFeature, decimal> GetFeatures(QAction action)
{
Point self = ((Maze)GetNewState(action)).self;
QSearch qsearch = new QSearch(this);
Maze simpleMaze = new Maze() { maze = maze, self = self, goal = goal, width = width, height = height, walls = walls.ToArray() };
QSearchResult bestPath = qsearch.AStar(simpleMaze);
List<Point> bestOppMoves = new List<Point>();
foreach (Point o in opponent)
{
bestOppMoves.Add(o);
bestOppMoves.Add(new Point(o.X + 1, o.Y));
bestOppMoves.Add(new Point(o.X - 1, o.Y));
bestOppMoves.Add(new Point(o.X, o.Y + 1));
bestOppMoves.Add(new Point(o.X, o.Y - 1));
}
foreach (Point o in bestOppMoves.ToArray())
{
bestOppMoves.Add(new Point(o.X + 1, o.Y));
bestOppMoves.Add(new Point(o.X - 1, o.Y));
bestOppMoves.Add(new Point(o.X, o.Y + 1));
bestOppMoves.Add(new Point(o.X, o.Y - 1));
}
Maze safeMaze = new Maze() { maze = maze, self = self, goal = goal, width = width, height = height, walls = walls.Concat(bestOppMoves).ToArray() };
QSearchResult safePath = qsearch.AStar(safeMaze);
Dictionary<string, decimal> features = new Dictionary<string, decimal>() {
//{this.GetHashCode().ToString()+"_"+action, 1}, // Identity to convert this back to QLearning
{"Goal", goal==self? 1:0},
{"Direct_Distance_To_Goal", bestPath==null? 1:(decimal) bestPath.Count / (decimal)(width * height)},
{"Safe_Distance_To_Goal", (safePath==null? 1: (decimal)safePath.Count/ (decimal)(width * height))}
};
decimal distanceToOpponent = decimal.MaxValue;
if (opponent.Any())
{
features["Opponent"] = goal!=self && opponent.Where(p => (Math.Abs(p.X - self.X) <= 1 && p.Y == self.Y) || (Math.Abs(p.Y - self.Y) <= 1 && p.X == self.X)).Any() ? 1 : 0;
distanceToOpponent = opponent.Select(o => qsearch.AStar(new Maze() { maze = maze, self = self, goal = o, width = width, height = height, walls = walls })).Select(x=>x==null? width*height:x.Count).Min();
features["Distance_To_Opponent"] = distanceToOpponent>=5? 1:distanceToOpponent / (decimal)(width * height);
if (goal != self)
{
Maze deadEnd = new Maze() { maze = maze, self = self, goal = goal, width = width, height = height, walls = walls.Concat(new Point[] { new Point(this.self.X - 1, this.self.Y), new Point(this.self.X + 1, this.self.Y), new Point(this.self.X, this.self.Y - 1), new Point(this.self.X, this.self.Y + 1) }).ToArray() };
QSearchResult deadPath = qsearch.Depth_First(deadEnd);
if (deadPath == null)
{
features["Dead_End"] = 1;
}
}
}
return QFeature_String.FromStringDecimalDictionary(features);
}
public override void Step()
{
if (!HideOutput && maze!=null)
{
try
{
maze.SetSelf(self.X, self.Y);
}
catch (Exception e)
{
WriteOutput("" + e, true);
}
}
if (opponent.Any() && goal!=self)
{
for (int i = 0; i < opponent.Count; i++)
{
QState tempState = new Maze() { maze = maze, self = opponent[i], goal = self, width = width, height = height, walls = walls};
QSearchResult opponentPath;
if (bestOpponentPath.ContainsKey(tempState)) opponentPath = bestOpponentPath[tempState];
else
{
opponentPath = new QSearch().AStar(tempState);
bestOpponentPath[tempState] = opponentPath;
}
if (opponentPath!=null && opponentPath.Any())
{
int newX = opponent[i].X, newY = opponent[i].Y;
QAction action = opponentPath.actionsList.First();
if (random.NextDouble() > opponentDifficulty)
{
QAction[] allActions = tempState.GetActions();
action = allActions.ElementAt(random.Next(allActions.Length));
}
// Translate decision to new coordinate
if (action.ToString() == "up") newY--;
else if (action.ToString() == "down") newY++;
else if (action.ToString() == "left") newX--;
else if (action.ToString() == "right") newX++;
//WriteOutput("Moving adversary at " + opponent[i] + " " + action + " to "+newX+", "+newY+"...");
opponent[i] = new Point(newX, newY);
if (!HideOutput && maze!=null)
{
maze.SetOpponent(i, newX, newY);
}
}
}
}
if (goal == self)
{
score += 100;
}
else if (opponent.Contains(self))
{
score -= 100;
}
if (!HideOutput && CurrentMode == AWAKEN) Thread.Sleep(100);
}