StateEvalScore UCTSearch::traverse(UCTNode & node, GameState & currentState)
{
StateEvalScore playoutVal;
_results.totalVisits++;
// if we haven't visited this node yet, do a playout
if (node.numVisits() == 0)
{
// update the status of the current state with this node's moves
//updateState(node, currentState, !node.hasChildren());
updateState(node, currentState, true);
// do the playout
playoutVal = currentState.eval(_params.maxPlayer(), _params.evalMethod(), _params.simScript(Players::Player_One), _params.simScript(Players::Player_Two));
_results.nodesVisited++;
}
// otherwise we have seen this node before
else
{
// update the state for a non-leaf node
updateState(node, currentState, false);
if (currentState.isTerminal())
{
playoutVal = currentState.eval(_params.maxPlayer(), EvaluationMethods::LTD2);
}
else
{
// if the children haven't been generated yet
if (!node.hasChildren())
{
generateChildren(node, currentState);
}
UCTNode & next = UCTNodeSelect(node);
playoutVal = traverse(next, currentState);
}
}
node.incVisits();
if (playoutVal.val() > 0)
{
node.addWins(1);
}
else if (playoutVal.val() == 0)
{
node.addWins(0.5);
}
return playoutVal;
}
void GreedyPlayer::run()
{
const QList<Move> moves = state->moves(type);
GameState next;
double bestEval;
if (type == Attack) bestEval = GameState::evalMin;
else bestEval = GameState::evalMax;
foreach (const Move& m, moves) {
state->copyAndMove(m, next);
double eval = next.eval();
if ((type == Attack && eval > bestEval) ||
(type == Defense && eval < bestEval)) {
bestEval = eval;
chosen_move = m;
}
}
AlphaBetaValue AlphaBeta::alphaBeta(GameState & state, size_t depth, const IDType lastPlayerToMove, const MoveTuple * prevSimMove, AlphaBetaScore alpha, AlphaBetaScore beta)
{
// update statistics
_results.nodesExpanded++;
if (searchTimeOut())
{
throw 1;
}
if (terminalState(state, depth))
{
// return the value, but the move will not be valid since none was performed
AlphaBetaScore evalScore = state.eval(_params.maxPlayer(), _params.evalMethod(), _params.modelSimMethod());
return AlphaBetaValue(AlphaBetaScore(evalScore.val(), state.getNumMovements(_params.maxPlayer()) + evalScore.numMoves() ), AlphaBetaMove(0, false));
}
// figure out which player is to move
const IDType playerToMove(getPlayerToMove(state, depth, lastPlayerToMove, !prevSimMove));
// is the player to move the max player?
bool maxPlayer = (playerToMove == _params.maxPlayer());
// Transposition Table Logic
TTLookupValue TTval;
if (isTranspositionLookupState(state, prevSimMove))
{
TTval = TTlookup(state, alpha, beta, depth);
// if this is a TT cut, return the proper value
if (TTval.cut())
{
return AlphaBetaValue(TTval.entry()->getScore(), getAlphaBetaMove(TTval, playerToMove));
}
}
bool bestMoveSet(false);
// move generation
MoveArray & moves = _allMoves[depth];
state.generateMoves(moves, playerToMove);
generateOrderedMoves(state, moves, TTval, playerToMove, depth);
// while we have more simultaneous move tuples
AlphaBetaMove bestMove, bestSimResponse;
MoveTuple numMoveTuples(getNumMoveTuples(moves, TTval, playerToMove, depth));
for (MoveTuple t(0); t < numMoveTuples; ++t)
{
// get the tuple that will be implemented
const MoveTuple tuple = getNextMoveTuple(t, depth);
// the value of the recursive AB we will call
AlphaBetaValue val;
// generate the child state
GameState child(state);
bool firstMove = true;
// if this is the first player in a simultaneous move state
if (state.bothCanMove() && !prevSimMove && (depth != 1))
{
firstMove = true;
// don't generate a child yet, just pass on the move we are investigating
val = alphaBeta(state, depth-1, playerToMove, &tuple, alpha, beta);
}
else
{
firstMove = false;
// if this is the 2nd move of a simultaneous move state
if (prevSimMove)
{
// do the previous move tuple selected by the first player to move during this state
doTupleMoves(child, _allMoves[depth+1], *prevSimMove);
}
// do the moves of the current player
doTupleMoves(child, moves, tuple);
child.finishedMoving(true);
// get the alpha beta value
val = alphaBeta(child, depth-1, playerToMove, NULL, alpha, beta);
}
// set alpha or beta based on maxplayer
if (maxPlayer && (val.score() > alpha))
{
alpha = val.score();
bestMove = AlphaBetaMove(tuple, true);
bestMoveSet = true;
if (state.bothCanMove() && !prevSimMove)
{
bestSimResponse = val.abMove();
}
// if this is depth 1 of the first try at depth 1, store the best in results
}
else if (!maxPlayer && (val.score() < beta))
{
beta = val.score();
bestMove = AlphaBetaMove(tuple, true);
bestMoveSet = true;
if (state.bothCanMove() && prevSimMove)
{
bestSimResponse = val.abMove();
}
}
if (alpha.val() == -10000000 && beta.val() == 10000000)
{
fprintf(stderr, "\n\nALPHA BETA ERROR, NO VALUE SET\n\n");
}
// alpha-beta cut
if (alpha >= beta)
{
break;
}
}
if (isTranspositionLookupState(state, prevSimMove))
{
TTsave(state, maxPlayer ? alpha : beta, alpha, beta, depth, playerToMove, bestMove, bestSimResponse);
}
return maxPlayer ? AlphaBetaValue(alpha, bestMove) : AlphaBetaValue(beta, bestMove);
}
