public Opt_SP_UCTTreeNode(IPuzzleMove move, Opt_SP_UCTTreeNode parent, IPuzzleState state, MersenneTwister rng, bool ucb1Tuned, bool rave, double raveThreshold, bool nodeRecycling, double const_C = 1, double const_D = 20000, bool generateUntriedMoves = true)
{
Move = move;
this.parent = parent;
this.const_C = const_C;
this.const_D = const_D;
rnd = rng;
childNodes = new List <Opt_SP_UCTTreeNode>();
NextLRUElem = null;
PrevLRUElem = null;
SetActive = false;
wins = 0;
visits = 0;
squares_rewards = 0;
RAVEwins = 0;
RAVEvisits = 0;
squaredReward = 0;
topScore = double.MinValue;
this.ucb1Tuned = ucb1Tuned;
this.rave = rave;
this.raveThreshold = raveThreshold;
this.nodeRecycling = nodeRecycling;
if (generateUntriedMoves)
{
untriedMoves = state.GetMoves();
}
}
public void DoMove(IPuzzleMove move)
{
stateChanged = false;
SamegameGameMove sgmove = move as SamegameGameMove;
//int value = board[sgmove.X][sgmove.Y];
//HashSet<Position> visited = new HashSet<Position>();
//List<Position> toRemove = new List<Position>();
//CheckAdjacentBlocks(sgmove.X, sgmove.Y, value, ref visited, ref toRemove); //remove adjacent blocks
if (sgmove.Blocks.Count > 0)
{
board[sgmove.X][sgmove.Y] = 1000;
score += (int)Math.Pow((sgmove.Blocks.Count() - 2), 2);
stateChanged = true;
}
foreach (Position position in sgmove.Blocks)
{
board[position.X][position.Y] = 1000;
}
for (int i = 0; i < board.Count; i++)
{
board[i].RemoveAll(v => v == 1000);
}
board.RemoveAll(column => column.Count == 0); //remove empty columns
}
private static void SokobanRandom(string levelPath, int maxDepth, int iterations, uint seed)
{
string[] levels = ReadSokobanLevels(levelPath);
int randomSolved = 0;
List <IPuzzleMove> moves = null;
RNG.Seed(seed + threadIndex);
MersenneTwister rng;
for (int i = 0; i < levels.Length; i++)
{
RNG.Seed(seed + threadIndex);
rng = new MersenneTwister(seed + threadIndex);
IPuzzleState state = new AbstractSokobanState(levels[i], RewardType.R0, useNormalizedPosition, useGoalMacro, useTunnelMacro, useGoalCut, null, rng);
IPuzzleMove move = null;
int restart = 0;
IPuzzleState clone = state.Clone();
string solution;
while (!state.EndState() && restart < iterations)
{
moves = new List <IPuzzleMove>();
state = clone.Clone();
int count = 0;
while (!state.isTerminal() && count < maxDepth)
{
move = state.GetRandomMove();
moves.Add(move);
state.DoMove(move);
count++;
}
restart++;
}
if (state.EndState())
{
solution = "";
int pushCount = 0;
foreach (SokobanPushMove push in moves)
{
foreach (IPuzzleMove m in push.MoveList)
{
if (m > 3)
{
pushCount++;
}
solution += m;
}
}
Log("Level " + (i + 1) + "\titerations: " + restart + "\tsolution length: (moves/pushes)" + solution.Length + "/" + pushCount + "\tsolution: " + solution);
randomSolved++;
}
else
{
Log("Level " + (i + 1) + "\tNo solution found");
}
}
Log("Solved: " + randomSolved + "/" + levels.Length);
}
public virtual ISPTreeNode AddChild(IPuzzleMove move, IPuzzleState state)
{
untriedMoves.Remove(move);
SP_UCTTreeNode n = new SP_UCTTreeNode(move, this, state, rnd, const_C, const_D, true);
childNodes.Add(n);
return(n);
}
public List <IPuzzleMove> Solve(IPuzzleState rootState, int iterations, double maxTimeInMinutes = 5)
{
topScore = double.MinValue;
bestRollout = null;
IPuzzleMove bestMove = Search(rootState, iterations, maxTimeInMinutes);
List <IPuzzleMove> moves = new List <IPuzzleMove>()
{
bestMove
};
moves.AddRange(bestRollout);
return(moves);
}
/// <summary>
/// Returns an active object from the object pool without resetting any of its values.
/// You will need to set its values and set it inactive again when you are done with it.
/// </summary>
/// <returns>ITreeNode of requested type if it is available, otherwise null.</returns>
public Opt_SP_UCTTreeNode GetObject(IPuzzleMove move, Opt_SP_UCTTreeNode parent, IPuzzleState state, MersenneTwister rng, bool ucb1Tuned, bool rave, double raveThreshold, bool nodeRecycling, double const_C, double const_D)
{
//iterate through all pooled objects.
foreach (Opt_SP_UCTTreeNode node in pooledObjects)
{
//look for the first one that is inactive.
if (node.SetActive)
{
continue;
}
//set the object to active.
node.SetActive = true;
//set object's values
node.Move = move;
node.Parent = parent;
node.untriedMoves = state.GetMoves();
node.Rnd = rng;
node.Ucb1Tuned = ucb1Tuned;
node.Rave = rave;
node.RaveThreshold = raveThreshold;
node.NodeRecycling = nodeRecycling;
node.ConstC = const_C;
node.ConstD = const_D;
//return the object we found.
return(node);
}
//if we make it this far, we obviously didn't find an inactive object.
//so we need to see if we can grow beyond our current count.
//if we reach the maximum size we didn't have any inactive objects.
//we also were unable to grow, so return null as we can't return an object.
if (maxPoolSize <= pooledObjects.Count)
{
return(null);
}
//Instantiate a new object.
//set it to active since we are about to use it.
Opt_SP_UCTTreeNode objNode = new Opt_SP_UCTTreeNode(move, parent, state, rng, ucb1Tuned, rave, raveThreshold, nodeRecycling, const_C, const_D)
{
SetActive = true
};
//add it to the pool of objects
pooledObjects.Add(objNode);
//return the object to the requestor.
return(objNode);
}
public IPuzzleMove selectMove(IPuzzleState gameState)
{
List <IPuzzleMove> moves = gameState.GetMoves();
if (rnd.NextDouble() <= 0.00007) //epsilon greedy
{
return(moves[rnd.Next(moves.Count)]);
}
moves.RemoveAll(item => gameState.GetBoard(SamegameGameMove.GetX(item), SamegameGameMove.GetY(item)) == selectedColor);
if (moves.Count == 0)
{
moves = gameState.GetMoves();
}
IPuzzleMove selectedMove = moves[rnd.Next(moves.Count)];
return(selectedMove);
}
public SP_UCTTreeNode(IPuzzleMove move, SP_UCTTreeNode parent, IPuzzleState state, MersenneTwister rng, double const_C = 1, double const_D = 20000, bool generateUntriedMoves = true)
{
this.move = move;
this.parent = parent;
this.const_C = const_C;
this.const_D = const_D;
rnd = rng;
childNodes = new List <SP_UCTTreeNode>();
wins = 0;
visits = 0;
squaredReward = 0;
topScore = double.MinValue;
if (generateUntriedMoves)
{
untriedMoves = state.GetMoves();
}
}
public IPuzzleMove selectMove(IPuzzleState gameState)
{
List <IPuzzleMove> moves = gameState.GetMoves();
IPuzzleMove bestMove = null;
if (rng.NextDouble() > epsilon)
{
IPuzzleState clone = gameState.Clone();
double maxReward = double.MinValue;
List <IPuzzleMove> bestMoves = new List <IPuzzleMove>();
foreach (IPuzzleMove move in clone.GetMoves())
{
clone.DoMove(move);
double result = clone.GetResult();
if (result > maxReward)
{
bestMoves.Clear();
bestMoves.Add(move);
maxReward = result;
//bestMove = move;
}
else if (result == maxReward)
{
bestMoves.Add(move);
}
clone = gameState.Clone();
}
//return bestMoves[0];
return(bestMoves[rng.Next(bestMoves.Count())]);
}
else
{
bestMove = gameState.GetRandomMove();
}
return(bestMove);
}
public AStarNode(IPuzzleState state, IPuzzleMove move, AStarNode parent)
{
this.state = state;
this.move = move;
this.parent = parent;
}
private static void ManualSokoban()
{
string level = " #####\n # ####\n # # #\n ## .#\n### ###.#\n# $ # #.#\n# $$# ###\n#@ #\n#####";
//string level = "####\n# .#\n# ###\n#*@ #\n# $ #\n# ###\n####";
Log("Level:\n" + level);
MersenneTwister rng = new MersenneTwister(1 + threadIndex);
ISPSimulationStrategy simulationStrategy = new SokobanRandomStrategy();
SokobanGameState s = new SokobanGameState(level, RewardType.NegativeBM, simulationStrategy);
SokobanGameState backupState = (SokobanGameState)s.Clone();
bool quit = false;
IPuzzleMove move = null;
Console.WriteLine(s.PrettyPrint());
while (!quit)
{
ConsoleKeyInfo input = Console.ReadKey();
List <IPuzzleMove> moves = s.GetMoves();
switch (input.Key)
{
case ConsoleKey.UpArrow:
if (moves.Contains(new SokobanGameMove("u")))
{
move = new SokobanGameMove("u");
}
else
{
move = new SokobanGameMove("U");
}
break;
case ConsoleKey.DownArrow:
if (moves.Contains(new SokobanGameMove("d")))
{
move = new SokobanGameMove("d");
}
else
{
move = new SokobanGameMove("D");
}
break;
case ConsoleKey.LeftArrow:
if (moves.Contains(new SokobanGameMove("l")))
{
move = new SokobanGameMove("l");
}
else
{
move = new SokobanGameMove("L");
}
break;
case ConsoleKey.RightArrow:
if (moves.Contains(new SokobanGameMove("r")))
{
move = new SokobanGameMove("r");
}
else
{
move = new SokobanGameMove("R");
}
break;
case ConsoleKey.R:
s = (SokobanGameState)backupState.Clone();
move = null;
break;
case ConsoleKey.Q:
move = null;
quit = true;
break;
}
if (move != null)
{
Console.WriteLine("Move: " + move);
s.DoMove(move);
}
Console.WriteLine(s.PrettyPrint());
Console.WriteLine("Score: " + s.GetScore() + " | isTerminal: " + s.isTerminal());
}
}
public BFSNodeState(IPuzzleState state, IPuzzleMove move, BFSNodeState parent)
{
this.state = state;
this.move = move;
this.parent = parent;
}
public ISPTreeNode AddChild(ObjectPool objectPool, IPuzzleMove move, IPuzzleState state)
{
throw new NotImplementedException();
}
public void RemoveUntriedMove(IPuzzleMove move)
{
untriedMoves.Remove(move);
}
public IPuzzleMove Search(IPuzzleState rootState, int iterations, double maxTimeInMinutes = 5)
{
IterationsForFirstSolution = -1;
nodeCount = 0;
nodesEliminated = 0;
nodesNotExpanded = 0;
bool looped;
if (!search)
{
search = true;
}
double minReward = double.MaxValue;
double maxReward = double.MinValue;
maxDepth = 0;
// If needed clean the pool, restore all objects in the pool to the initial value
if (objectPool.NeedToClean)
{
objectPool.CleanObjectPool();
}
ISPTreeNode rootNode = treeCreator.GenRootNode(rootState);
ISPTreeNode head = null;
ISPTreeNode tail = null;
HashSet <IPuzzleMove> allFirstMoves = new HashSet <IPuzzleMove>();
List <IPuzzleMove> currentRollout = new List <IPuzzleMove>();
solutionHashes = new HashSet <int>();
#if PROFILING
long beforeMemory = GC.GetTotalMemory(false);
long afterMemory = GC.GetTotalMemory(false);
long usedMemory = afterMemory - beforeMemory;
long averageUsedMemoryPerIteration = 0;
#endif
int deadlocksInTree = 0;
int currentDepth = 0;
for (iterationsExecuted = 0; iterationsExecuted < iterations; iterationsExecuted++)
{
looped = false;
ISPTreeNode node = rootNode;
IPuzzleState state = rootState.Clone();
//Debug.WriteLine(node.TreeToString(0));
HashSet <IPuzzleState> visitedStatesInRollout = new HashSet <IPuzzleState>()
{
state.Clone()
};
// Clear lists of moves used for RAVE updates && best rollout
solutionHash = 27;
currentRollout = new List <IPuzzleMove>();
allFirstMoves.Clear();
// Select
while (!node.HasMovesToTry() && node.HasChildren())
{
// UCB1-Tuned and RAVE Optimizations
node = node.SelectChild();
state.DoMove(node.Move);
visitedStatesInRollout.Add(state.Clone());
// RAVE Optimization && best rollout
currentRollout.Add(node.Move);
UpdateSolutionHash(node.Move);
allFirstMoves.Add(node.Move);
// Node Recycling Optimization
if (((Opt_SP_UCTTreeNode)node).NodeRecycling)
{
// Non-leaf node removed from LRU queue during playout
if (node.NextLRUElem != null && node.PrevLRUElem != null)
{
LRUQueueManager.LRURemoveElement(ref node, ref head, ref tail);
}
}
}
IPuzzleState backupState = state.Clone();
if (!node.HasChildren() && !node.HasMovesToTry())
{
deadlocksInTree++;
}
else
{
Debug.Write("");
}
// Expand
if (node.HasMovesToTry())
{
IPuzzleMove move = node.SelectUntriedMove();
if (move != -1)
{
state.DoMove(move);
// Node Recycling Optimization
if (((Opt_SP_UCTTreeNode)node).NodeRecycling)
{
if (memoryBudget == nodeCount && head != null)
{
head.ChildRecycle();
nodeCount--;
// Change LRU queue head when it becomes a leaf node
if (!head.HasChildren())
{
LRUQueueManager.LRURemoveFirst(ref head, ref tail);
}
}
}
if (visitedStatesInRollout.Contains(state))
{
if (avoidCycles)
{
while (node.GetUntriedMoves().Count > 0 && visitedStatesInRollout.Contains(state))
{
state = backupState.Clone();
move = node.GetUntriedMoves()[RNG.Next(node.GetUntriedMoves().Count)];
state.DoMove(move);
node.RemoveUntriedMove(move);
}
if (!visitedStatesInRollout.Contains(state)) //found valid move
{
node = node.AddChild(objectPool, move, state);
UpdateSolutionHash(move);
currentRollout.Add(move);
allFirstMoves.Add(move);
nodeCount++;
}
else //all moves visited
{
nodesNotExpanded++;
state = backupState;
}
}
else
{
nodesNotExpanded++;
looped = true;
}
}
else
{
node = node.AddChild(objectPool, move, state);
// RAVE Optimization && best rollout
UpdateSolutionHash(move);
currentRollout.Add(move);
allFirstMoves.Add(move);
nodeCount++;
}
visitedStatesInRollout.Add(state.Clone());
}
else
{
state.Pass();
}
}
else
{
nodesNotExpanded++;
}
// Rollout
while (!state.isTerminal() && !looped)
{
var move = state.GetSimulationMove();
backupState = state.Clone();
if (move != -1)
{
state.DoMove(move);
if (visitedStatesInRollout.Contains(state))
{
if (avoidCycles)
{
state = backupState.Clone();
List <IPuzzleMove> availableMoves = state.GetMoves();
while (availableMoves.Count > 0 && visitedStatesInRollout.Contains(state))
{ //keep trying different moves until we end up in an unvisited state
state = backupState.Clone();
move = availableMoves[RNG.Next(availableMoves.Count)];
availableMoves.Remove(move);
state.DoMove(move);
}
if (availableMoves.Count == 0 && visitedStatesInRollout.Contains(state))//all states have already been visited
{
break;
}
}
else
{
looped = true;
}
}
// RAVE Optimization && best rollout
UpdateSolutionHash(move);
currentRollout.Add(move);
allFirstMoves.Add(move);
visitedStatesInRollout.Add(state.Clone());
}
else //simulation ended
{
break;
//state.Pass();
}
}
//Keep topScore and update bestRollout
double result = state.GetResult();
minReward = Math.Min(result, minReward);
maxReward = Math.Max(result, maxReward);
if (state.EndState() && !solutionHashes.Contains(solutionHash))
{
solutionHashes.Add(solutionHash);
solutionCount++;
if (iterationsForFirstSolution < 0)
{
iterationsForFirstSolution = iterationsExecuted + 1;
}
}
if (result > topScore || result == topScore && currentRollout.Count < bestRollout.Count)
{
topScore = result;
bestRollout = currentRollout;
if (state.EndState() && stopOnResult)
{
iterationsExecuted++;
break;
}
}
// Backpropagate
currentDepth = 0;
while (node != null)
{
if (looped)
{
//TODO penalize score for loops?
}
ISPTreeNode parent = node.Parent;
//if a node is a dead end remove it from the tree
if (!node.HasChildren() && !node.HasMovesToTry() && !state.EndState() && useNodeElimination)
{
if (node.Parent == null)//unsolvable level. The tree has been completely explored. Return current best score
{
//SinglePlayerMCTSMain.Log("Unsolvable Level");
//Console.WriteLine("\nUnsolvable Level");
break;
}
node.Parent.RemoveChild(node);
nodeCount--;
nodesEliminated++;
currentDepth--;
}
// RAVE Optimization
node.Update(result, allFirstMoves);
node = parent;
currentDepth++;
// Node Recycling Optimization
if (((Opt_SP_UCTTreeNode)rootNode).NodeRecycling)
{
// Non-leaf node pushed back to LRU queue when updated
if (node != rootNode && node != null && node.HasChildren())
{
LRUQueueManager.LRUAddLast(ref node, ref head, ref tail);
}
}
}
maxDepth = Math.Max(maxDepth, currentDepth);
if (!rootNode.HasChildren() && !rootNode.HasMovesToTry())
{
break;
}
if (!search)
{
search = true;
return(null);
}
#if PROFILING
afterMemory = GC.GetTotalMemory(false);
usedMemory = afterMemory - beforeMemory;
averageUsedMemoryPerIteration = usedMemory / (i + 1);
var outStringToWrite = string.Format(" optMCTS search: {0:0.00}% [{1} of {2}] - Total used memory B(MB): {3}({4:N7}) - Average used memory per iteration B(MB): {5}({6:N7})\n",
(float)((i + 1) * 100) / (float)iterations, i + 1, iterations, usedMemory, usedMemory / 1024 / 1024, averageUsedMemoryPerIteration,
(float)averageUsedMemoryPerIteration / 1024 / 1024);
#if DEBUG
if (showMemoryUsage)
{
Console.Write(outStringToWrite);
Console.SetCursorPosition(0, Console.CursorTop);
}
#endif
#endif
//Console.WriteLine(rootNode.TreeToString(0));
}
//Console.WriteLine();
objectPool.NeedToClean = true;
//#if DEBUG
// Console.WriteLine(rootNode.ChildrenToString());
// Console.WriteLine(rootNode.TreeToString(0));
//#endif
IPuzzleMove bestMove;
if (bestRollout != null && bestRollout.Count > 0) //Remove first move from rollout so that if the topScore is not beaten we can just take the next move on the next search
{
bestMove = bestRollout[0];
bestRollout.RemoveAt(0);
}
else
{
bestMove = rootNode.GetBestMove();
}
Debug.WriteLine(rootNode.TreeToString(0));
Debug.WriteLine("Min Reward: " + minReward + " - Max Reward: " + maxReward);
visits = new List <int>();
raveVisits = new List <int>();
CountVisits((Opt_SP_UCTTreeNode)rootNode, visits, raveVisits);
visits.Sort((x, y) => (x.CompareTo(y)));
raveVisits.Sort((x, y) => (x.CompareTo(y)));
//string visitsString = LogVisits((Opt_SP_UCTTreeNode) rootNode);
//SinglePlayerMCTSMain.Log("Iterations: "+IterationsExecuted+" NodeCount: " + nodeCount+" "+visitsString);
return(bestMove);
}