tt: 修复和未开启置换表的对战，执黑胜率少3%，执白胜率少6%的问题
同时对于 alpha-beta 的函数传参形式进行修改： search(depth - 1 + epsilon, alpha, beta, node->children[i]); 改为 -search(depth - 1 + epsilon, -beta, -alpha, node->children[i]); 的形式。 简化必败以及明显劣势的判断方式。只要判断 root 的值，不需要看其孩子。 并顺带： 删除 MIN_MAX_ONLY 宏相关代码； 删除 SORT_CONSIDER_PRUNED、CLEAR_PRUNED_FLAG_BEFORE_SEARCH 宏相关代码； 新的 TT 算法和修改前算法对战 6000 盘。 黑 22% : 10%， 白 86% : 74% 有 12% 的优势。 速度提升 1.5%, 可以认为无变化。 和未开TT对战1700盘，均为 11% : 83%，误差不超过1%，故可认为新TT已经不会 造成明显劣化。开TT和未开TT比，目前是4.6倍速度比。 自对弈棋谱会变化。自对弈时长 12s+, 最后一步是 -(1,8), 白方胜。
This commit is contained in:
Calcitem
parent
1f62f4469f
commit
7b566b1e71
|
|
@ -62,8 +62,6 @@
|
|||
//#define TIME_STAT
|
||||
//#define CYCLE_STAT
|
||||
|
||||
//#define MIN_MAX_ONLY
|
||||
|
||||
//#define EVALUATE_ENABLE
|
||||
|
||||
#ifdef EVALUATE_ENABLE
|
||||
|
|
@ -82,7 +80,6 @@
|
|||
//#define IDS_DEBUG
|
||||
//#define IDS_ADD_VALUE
|
||||
|
||||
//#define CLEAR_PRUNED_FLAG_BEFORE_SEARCH
|
||||
//#define DEEPER_IF_ONLY_ONE_LEGAL_MOVE
|
||||
|
||||
#define TRANSPOSITION_TABLE_ENABLE
|
||||
|
|
@ -109,8 +106,6 @@
|
|||
|
||||
//#define DONOT_DELETE_TREE
|
||||
|
||||
//#define SORT_CONSIDER_PRUNED
|
||||
|
||||
//#define MESSAGEBOX_ENABLE
|
||||
|
||||
#ifdef DEBUG_MODE
|
||||
|
|
@ -153,4 +148,4 @@
|
|||
#define likely(expr) (__builtin_expect(!!(expr), 1))
|
||||
#define unlikely(expr) (__builtin_expect(!!(expr), 0))
|
||||
|
||||
#endif // CONFIG_H
|
||||
#endif // CONFIG_H
|
||||
|
|
@ -140,6 +140,10 @@ value_t Evaluation::getValue(StateInfo *st, Position *position, Node *node)
|
|||
break;
|
||||
}
|
||||
|
||||
if (st->position->sideToMove == PLAYER_WHITE) {
|
||||
value = -value;
|
||||
}
|
||||
|
||||
// 赋值返回
|
||||
node->value = value;
|
||||
return value;
|
||||
|
|
|
|||
|
|
@ -207,10 +207,6 @@ Node *Node::addChild(
|
|||
#ifdef ALPHABETA_AI
|
||||
newNode->value = VALUE_ZERO;
|
||||
newNode->rating = RATING_ZERO;
|
||||
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
newNode->pruned = false;
|
||||
#endif
|
||||
#endif // ALPHABETA_AI
|
||||
|
||||
newNode->childrenSize = 0; // Important
|
||||
|
|
@ -395,9 +391,7 @@ int AIAlgorithm::nodeCompare(const Node *first, const Node *second)
|
|||
return 0;
|
||||
}
|
||||
|
||||
int ret = (gSideToMove == PLAYER_BLACK ? 1 : -1);
|
||||
|
||||
return (first->value < second->value ? ret : -ret);
|
||||
return (first->value < second->value ? 1 : -1);
|
||||
}
|
||||
|
||||
return (first->rating < second->rating ? 1 : -1);
|
||||
|
|
@ -411,9 +405,9 @@ void AIAlgorithm::sortMoves(Node *node)
|
|||
|
||||
//#define DEBUG_SORT
|
||||
#ifdef DEBUG_SORT
|
||||
for (int i = 0; i < node->childrenSize; i++) {
|
||||
for (int moveIndex = 0; moveIndex < node->childrenSize; moveIndex++) {
|
||||
loggerDebug("* [%d] %p: %d = %d %d (%d)\n",
|
||||
i, &(node->children[i]), node->children[i]->move, node->children[i]->value, node->children[i]->rating, !node->children[i]->pruned);
|
||||
moveIndex, &(node->children[moveIndex]), node->children[moveIndex]->move, node->children[moveIndex]->value, node->children[moveIndex]->rating, !node->children[moveIndex]->pruned);
|
||||
}
|
||||
loggerDebug("\n");
|
||||
#endif
|
||||
|
|
@ -454,14 +448,14 @@ void AIAlgorithm::sortMoves(Node *node)
|
|||
|
||||
#ifdef DEBUG_SORT
|
||||
if (st->position.sideToMove == PLAYER_BLACK) {
|
||||
for (int i = 0; i < node->childrenSize; i++) {
|
||||
for (int moveIndex = 0; moveIndex < node->childrenSize; moveIndex++) {
|
||||
loggerDebug("+ [%d] %p: %d = %d %d (%d)\n",
|
||||
i, &(node->children[i]), node->children[i]->move, node->children[i]->value, node->children[i]->rating, !node->children[i]->pruned);
|
||||
moveIndex, &(node->children[moveIndex]), node->children[moveIndex]->move, node->children[moveIndex]->value, node->children[moveIndex]->rating, !node->children[moveIndex]->pruned);
|
||||
}
|
||||
} else {
|
||||
for (int i = 0; i < node->childrenSize; i++) {
|
||||
for (int moveIndex = 0; moveIndex < node->childrenSize; moveIndex++) {
|
||||
loggerDebug("- [%d] %p: %d = %d %d (%d)\n",
|
||||
i, &(node->children[i]), node->children[i]->move, node->children[i]->value, node->children[i]->rating, !node->children[i]->pruned);
|
||||
moveIndex, &(node->children[moveIndex]), node->children[moveIndex]->move, node->children[moveIndex]->value, node->children[moveIndex]->rating, !node->children[moveIndex]->pruned);
|
||||
}
|
||||
}
|
||||
loggerDebug("\n----------------------------------------\n");
|
||||
|
|
@ -592,26 +586,22 @@ int AIAlgorithm::search(depth_t depth)
|
|||
loggerDebug("%d(%d) ", value, value - lastValue);
|
||||
|
||||
#ifdef IDS_DEBUG
|
||||
loggerDebug(": --------------- depth = %d/%d ---------------\n", i, d);
|
||||
loggerDebug(": --------------- depth = %d/%d ---------------\n", moveIndex, d);
|
||||
int k = 0;
|
||||
int cs = root->childrenSize;
|
||||
for (int j = 0; j < cs; j++) {
|
||||
if (root->children[j]->value == root->value
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
&& !root->children[j]->pruned
|
||||
#endif
|
||||
) {
|
||||
for (int i = 0; i < cs; i++) {
|
||||
if (root->children[i]->value == root->value) {
|
||||
loggerDebug("[%.2d] %d\t%s\t%d\t%d *\n", k,
|
||||
root->children[j]->move,
|
||||
moveToCommand(root->children[j]->move),
|
||||
root->children[j]->value,
|
||||
root->children[j]->rating);
|
||||
root->children[i]->move,
|
||||
moveToCommand(root->children[i]->move),
|
||||
root->children[i]->value,
|
||||
root->children[i]->rating);
|
||||
} else {
|
||||
loggerDebug("[%.2d] %d\t%s\t%d\t%d\n", k,
|
||||
root->children[j]->move,
|
||||
moveToCommand(root->children[j]->move),
|
||||
root->children[j]->value,
|
||||
root->children[j]->rating);
|
||||
root->children[i]->move,
|
||||
moveToCommand(root->children[i]->move),
|
||||
root->children[i]->value,
|
||||
root->children[i]->rating);
|
||||
}
|
||||
|
||||
k++;
|
||||
|
|
@ -663,6 +653,8 @@ int AIAlgorithm::search(depth_t depth)
|
|||
#endif // IDS_WINDOW
|
||||
}
|
||||
|
||||
originDepth = d;
|
||||
|
||||
value = search(d, alpha, beta, root);
|
||||
|
||||
#ifdef TIME_STAT
|
||||
|
|
@ -682,9 +674,6 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
// 评价值
|
||||
value_t value;
|
||||
|
||||
// 当前节点的 MinMax 值，最终赋值给节点 value，与 alpha 和 Beta 不同
|
||||
value_t minMax;
|
||||
|
||||
// 临时增加的深度，克服水平线效应用
|
||||
depth_t epsilon = 0;
|
||||
|
||||
|
|
@ -749,12 +738,6 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
#endif
|
||||
node->value = probeVal;
|
||||
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
if (type != TT::hashfEXACT && type != TT::hashfEMPTY) {
|
||||
node->pruned = true; // TODO: 是否有用?
|
||||
}
|
||||
#endif
|
||||
|
||||
#if 0
|
||||
// TODO: 有必要针对深度微调 value?
|
||||
if (position->turn == PLAYER_BLACK)
|
||||
|
|
@ -842,6 +825,7 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
);
|
||||
|
||||
if (node == root && moveSize == 1) {
|
||||
best = moves[0];
|
||||
return node->value;
|
||||
}
|
||||
}
|
||||
|
|
@ -849,18 +833,8 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
// 排序子节点树
|
||||
sortMoves(node);
|
||||
|
||||
// 根据演算模型执行 MiniMax 检索，对先手，搜索 Max, 对后手，搜索 Min
|
||||
|
||||
minMax = st->position->sideToMove == PLAYER_BLACK ? -VALUE_INFINITE : VALUE_INFINITE;
|
||||
|
||||
assert(node->childrenSize != 0);
|
||||
|
||||
#ifdef CLEAR_PRUNED_FLAG_BEFORE_SEARCH
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
node->pruned = false;
|
||||
#endif // SORT_CONSIDER_PRUNED
|
||||
#endif // CLEAR_PRUNED_FLAG_BEFORE_SEARCH
|
||||
|
||||
int nchild = node->childrenSize;
|
||||
|
||||
#ifdef TRANSPOSITION_TABLE_ENABLE
|
||||
|
|
@ -881,8 +855,10 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
for (int i = 0; i < nchild; i++) {
|
||||
// 棋局入栈保存，以便后续撤销着法
|
||||
stashPosition();
|
||||
|
||||
doMove(node->children[i]->move);
|
||||
player_t before = st->position->sideToMove;
|
||||
move_t m = node->children[i]->move;
|
||||
doMove(m);
|
||||
player_t after = st->position->sideToMove;
|
||||
|
||||
if (gameOptions.getDepthExtension() == true && nchild == 1) {
|
||||
epsilon = 1;
|
||||
|
|
@ -891,74 +867,37 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
}
|
||||
|
||||
// 递归 Alpha-Beta 剪枝
|
||||
value = search(depth - 1 + epsilon, alpha, beta, node->children[i]);
|
||||
if (after != before) {
|
||||
value = -search(depth - 1 + epsilon, -beta, -alpha, node->children[i]);
|
||||
} else {
|
||||
value = search(depth - 1 + epsilon, alpha, beta, node->children[i]);
|
||||
}
|
||||
|
||||
undoMove();
|
||||
|
||||
switch (st->position->sideToMove) {
|
||||
case PLAYER_BLACK:
|
||||
// 为走棋一方的层, 局面对走棋的一方来说是以 α 为评价
|
||||
|
||||
// 取最大值
|
||||
minMax = std::max(value, minMax);
|
||||
|
||||
// α 为走棋一方搜索到的最好值，任何比它小的值对当前结点的走棋方都没有意义
|
||||
// 如果某个着法的结果小于或等于 α，那么它就是很差的着法，因此可以抛弃
|
||||
|
||||
if (value > alpha) {
|
||||
if (value >= beta) {
|
||||
#ifdef TRANSPOSITION_TABLE_ENABLE
|
||||
hashf = TT::hashfEXACT;
|
||||
hashf = TT::hashfBETA;
|
||||
#endif
|
||||
alpha = value;
|
||||
}
|
||||
node->value = beta;
|
||||
goto out;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case PLAYER_WHITE:
|
||||
// 为走棋方的对手一方的层, 局面对对手一方来说是以 β 为评价
|
||||
|
||||
// 取最小值
|
||||
minMax = std::min(value, minMax);
|
||||
|
||||
// β 表示对手目前的劣势，这是对手所能承受的最坏结果
|
||||
// β 值越大，表示对手劣势越明显
|
||||
// 在对手看来，他总是会找到一个对策不比 β 更坏的
|
||||
// 如果当前结点返回 β 或比 β 更好的值，作为父结点的对方就绝对不会选择这种策略，
|
||||
// 如果搜索过程中返回 β 或比 β 更好的值，那就够好的了，走棋的一方就没有机会使用这种策略了。
|
||||
// 如果某个着法的结果大于或等于 β，那么整个结点就作废了，因为对手不希望走到这个局面，而它有别的着法可以避免到达这个局面。
|
||||
// 因此如果我们找到的评价大于或等于β，就证明了这个结点是不会发生的，因此剩下的合理着法没有必要再搜索。
|
||||
|
||||
// TODO: 本意是要删掉这句，忘了删，结果反而棋力没有明显问题，待查
|
||||
// 如果删掉这句，启用下面这段代码，则三有时不会堵并且计算效率较低
|
||||
// 有了这句之后，hashf 不可能等于 hashfBETA
|
||||
beta = std::min(value, beta);
|
||||
|
||||
#if 0
|
||||
if (value < beta) {
|
||||
if (value > alpha) {
|
||||
#ifdef TRANSPOSITION_TABLE_ENABLE
|
||||
hashf = hashfBETA;
|
||||
hashf = TT::hashfEXACT;
|
||||
#endif
|
||||
beta = value;
|
||||
alpha = value;
|
||||
|
||||
if (depth == originDepth) {
|
||||
best = m;
|
||||
}
|
||||
#endif
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
#ifndef MIN_MAX_ONLY
|
||||
// 如果某个着法的结果大于 α 但小于β，那么这个着法就是走棋一方可以考虑走的
|
||||
// 否则剪枝返回
|
||||
if (alpha >= beta) {
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
node->pruned = true;
|
||||
#endif
|
||||
break;
|
||||
}
|
||||
#endif /* !MIN_MAX_ONLY */
|
||||
}
|
||||
|
||||
node->value = minMax;
|
||||
node->value = alpha;
|
||||
|
||||
out:
|
||||
|
||||
#ifdef DEBUG_AB_TREE
|
||||
node->alpha = alpha;
|
||||
|
|
@ -978,15 +917,12 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
}
|
||||
#endif // DONOT_DELETE_TREE
|
||||
|
||||
|
||||
|
||||
if (gameOptions.getIDSEnabled()) {
|
||||
#ifdef IDS_ADD_VALUE
|
||||
if (st->position->sideToMove == PLAYER_BLACK) {
|
||||
node->children[0]->value += 1;
|
||||
node->value += 1;
|
||||
} else {
|
||||
node->children[0]->value -= 1;
|
||||
node->value -= 1;
|
||||
}
|
||||
node->children[0]->value += 1;
|
||||
node->value += 1;
|
||||
#endif /* IDS_ADD_VALUE */
|
||||
}
|
||||
|
||||
|
|
@ -997,7 +933,7 @@ value_t AIAlgorithm::search(depth_t depth, value_t alpha, value_t beta, Node *no
|
|||
hashf,
|
||||
hash
|
||||
#ifdef BEST_MOVE_ENABLE
|
||||
, node->children[0]->move
|
||||
, best
|
||||
#endif // BEST_MOVE_ENABLE
|
||||
);
|
||||
#endif /* TRANSPOSITION_TABLE_ENABLE */
|
||||
|
|
@ -1030,8 +966,7 @@ void AIAlgorithm::undoMove()
|
|||
#ifdef ALPHABETA_AI
|
||||
const char* AIAlgorithm::bestMove()
|
||||
{
|
||||
vector<Node*> bestMoves;
|
||||
size_t bestMovesSize = 0;
|
||||
char charChoose = '*';
|
||||
|
||||
if (!root->childrenSize) {
|
||||
return "error!";
|
||||
|
|
@ -1039,21 +974,22 @@ const char* AIAlgorithm::bestMove()
|
|||
|
||||
Board::printBoard();
|
||||
|
||||
int i = 0;
|
||||
int moveIndex = 0;
|
||||
|
||||
int cs = root->childrenSize;
|
||||
for (int j = 0; j < cs; j++) {
|
||||
if (root->children[j]->value == root->value
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
&& !root->children[j]->pruned
|
||||
#endif
|
||||
) {
|
||||
loggerDebug("[%.2d] %d\t%s\t%d\t%d *\n", i, root->children[j]->move, moveToCommand(root->children[j]->move), root->children[j]->value, root->children[j]->rating);
|
||||
} else {
|
||||
loggerDebug("[%.2d] %d\t%s\t%d\t%d\n", i, root->children[j]->move, moveToCommand(root->children[j]->move), root->children[j]->value, root->children[j]->rating);
|
||||
for (int i = 0; i < cs; i++) {
|
||||
if (root->children[i]->move != best) {
|
||||
charChoose = ' ';
|
||||
}
|
||||
|
||||
i++;
|
||||
loggerDebug("[%.2d] %d\t%s\t%d\t%d %c\n", moveIndex,
|
||||
root->children[i]->move,
|
||||
moveToCommand(root->children[i]->move),
|
||||
root->children[i]->value,
|
||||
root->children[i]->rating,
|
||||
charChoose);
|
||||
|
||||
moveIndex++;
|
||||
}
|
||||
|
||||
player_t side = state->position->sideToMove;
|
||||
|
|
@ -1061,17 +997,7 @@ const char* AIAlgorithm::bestMove()
|
|||
#ifdef ENDGAME_LEARNING
|
||||
// 检查是否明显劣势
|
||||
if (gameOptions.getLearnEndgameEnabled()) {
|
||||
bool isMostWeak = true; // 是否明显劣势
|
||||
|
||||
for (int j = 0; j < root->childrenSize; j++) {
|
||||
if ((side == PLAYER_BLACK && root->children[j]->value > -VALUE_STRONG) ||
|
||||
(side == PLAYER_WHITE && root->children[j]->value < VALUE_STRONG)) {
|
||||
isMostWeak = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (isMostWeak) {
|
||||
if (root->value <= -VALUE_STRONG) {
|
||||
Endgame endgame;
|
||||
endgame.type = state->position->sideToMove == PLAYER_BLACK ?
|
||||
ENDGAME_PLAYER_WHITE_WIN : ENDGAME_PLAYER_BLACK_WIN;
|
||||
|
|
@ -1083,40 +1009,13 @@ const char* AIAlgorithm::bestMove()
|
|||
|
||||
// 检查是否必败
|
||||
if (gameOptions.getGiveUpIfMostLose() == true) {
|
||||
bool isMostLose = true; // 是否必败
|
||||
|
||||
for (int j = 0; j < root->childrenSize; j++) {
|
||||
if ((side == PLAYER_BLACK && root->children[j]->value > -VALUE_WIN) ||
|
||||
(side == PLAYER_WHITE && root->children[j]->value < VALUE_WIN)) {
|
||||
isMostLose = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// 自动认输
|
||||
if (isMostLose) {
|
||||
if (root->value <= -VALUE_WIN) {
|
||||
sprintf(cmdline, "Player%d give up!", state->position->sideId);
|
||||
return cmdline;
|
||||
}
|
||||
}
|
||||
|
||||
int nchild = root->childrenSize;
|
||||
for (int j = 0; j < nchild; j++) {
|
||||
if (root->children[j]->value == root->value) {
|
||||
bestMoves.push_back(root->children[j]);
|
||||
}
|
||||
}
|
||||
|
||||
bestMovesSize = bestMoves.size();
|
||||
|
||||
if (bestMovesSize == 0) {
|
||||
loggerDebug("Not any child value is equal to root value\n");
|
||||
|
||||
for (int j = 0; j < root->childrenSize; j++) {
|
||||
bestMoves.push_back(root->children[j]);
|
||||
}
|
||||
}
|
||||
|
||||
loggerDebug("Evaluated: %llu / %llu = %llu%%\n", evaluatedNodeCount, nodeCount, evaluatedNodeCount * 100 / nodeCount);
|
||||
memmgr.memmgr_print_stats();
|
||||
|
||||
|
|
@ -1134,11 +1033,7 @@ const char* AIAlgorithm::bestMove()
|
|||
#endif // TRANSPOSITION_TABLE_DEBUG
|
||||
#endif // TRANSPOSITION_TABLE_ENABLE
|
||||
|
||||
if (bestMoves.empty()) {
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
return moveToCommand(bestMoves[0]->move);
|
||||
return moveToCommand(best);
|
||||
}
|
||||
#endif // ALPHABETA_AI
|
||||
|
||||
|
|
@ -1200,4 +1095,4 @@ void AIAlgorithm::loadEndgameFileToHashMap()
|
|||
const QString filename = "endgame.txt";
|
||||
endgameHashMap.load(filename);
|
||||
}
|
||||
#endif // ENDGAME_LEARNING
|
||||
#endif // ENDGAME_LEARNING
|
||||
|
|
@ -116,9 +116,6 @@ public:
|
|||
int childrenSize { 0 };
|
||||
|
||||
#ifdef ALPHABETA_AI
|
||||
#ifdef SORT_CONSIDER_PRUNED
|
||||
bool pruned { false };
|
||||
#endif
|
||||
value_t value { VALUE_UNKNOWN };
|
||||
rating_t rating { RATING_ZERO };
|
||||
#endif // ALPHABETA_AI
|
||||
|
|
@ -300,6 +297,9 @@ private:
|
|||
// 标识，用于跳出剪枝算法，立即返回
|
||||
bool requiredQuit {false};
|
||||
|
||||
move_t best { MOVE_NONE };
|
||||
depth_t originDepth { 0 };
|
||||
|
||||
private:
|
||||
// 命令行
|
||||
char cmdline[64] {};
|
||||
|
|
|
|||
Loading…
Reference in New Issue