int search_uct(int color, int node_n)
{
NODE *pN = &node[node_n];
CHILD *c = NULL;
int select, z, err, win, current_depth;
for (;;) {
select = select_best_ucb(node_n, color);
c = &pN->child[select];
z = c->z;
err = put_stone(z, color, FILL_EYE_ERR);
if ( err == 0 ) break;
c->z = ILLEGAL_Z; // select other move
}
current_depth = depth;
path[depth++] = c->z;
// playout in first time. <= 10 can reduce node.
if ( c->games <= 0 || depth == D_MAX || (c->z == 0 && depth>=2 && path[depth-2]==0) ) {
win = -playout(flip_color(color));
} else {
if ( c->next == NODE_EMPTY ) c->next = create_node(c->z);
win = -search_uct(flip_color(color), c->next);
}
update_rave(pN, color, current_depth, win);
// update winrate
c->rate = (c->rate * c->games + win) / (c->games + 1);
c->games++;
pN->child_games_sum++;
return win;
}
void UCTParallel::search_uct_root(Board& board, const Color color, UCTNode* node, UCTNode* copychild)
{
// UCBからプレイアウトする手を選択
// rootノードはアトミックに更新するためUCB計算ではロックしない
UCTNode* selected_node = select_node_with_ucb(node);
// rootでは全て合法手なのでエラーチェックはしない
board.move_legal(selected_node->xy, color);
// コピーされたノードに変換
UCTNode* selected_node_copy = copychild + (selected_node - node->child);
int win;
// 閾値以下の場合プレイアウト(全スレッドの合計値)
if (selected_node->playout_num < THR)
{
win = 1 - playout(board, opponent(color));
}
else {
if (selected_node_copy->child_num == 0)
{
// ノードを展開
if (selected_node_copy->expand_node(board))
{
win = 1 - search_uct(board, opponent(color), selected_node_copy);
}
else {
// ノードプール不足
win = 1 - playout(board, opponent(color));
}
}
else {
win = 1 - search_uct(board, opponent(color), selected_node_copy);
}
}
// 勝率を更新(アトミックに加算)
_InterlockedExchangeAdd(&selected_node->win_num, win);
_InterlockedIncrement(&selected_node->playout_num);
_InterlockedIncrement(&node->playout_num_sum);
}
int get_best_uct(int color){
int bestPlace=-1;
for (int i = 0; i < loop; i++){
int BoardCopy[BOARD_MAX];
int KoPlaceCopy = ko_z;
memcpy(BoardCopy,board,sizeof(board));
search_uct(color);
}
return bestPlace;
}
int get_best_uct(int color)
{
int next, i, best_z, best_i = -1;
int max = -999;
NODE *pN;
int prev_z = 0;
if ( moves > 0 ) prev_z = record[moves-1];
node_num = 0;
next = create_node(prev_z);
for (i=0; i<uct_loop; i++) {
int board_copy[BOARD_MAX];
int ko_z_copy = ko_z;
memcpy(board_copy, board, sizeof(board));
depth = 0;
search_uct(color, next);
ko_z = ko_z_copy;
memcpy(board, board_copy, sizeof(board));
}
pN = &node[next];
for (i=0; i<pN->child_num; i++) {
CHILD *c = &pN->child[i];
if ( c->games > max ) {
best_i = i;
max = c->games;
}
prt("%2d:z=%2d,rate=%6.3f,games=%4d, rave_r=%6.3f,g=%4d\n",
i, get81(c->z), c->rate, c->games, c->rave_rate, c->rave_games);
}
best_z = pN->child[best_i].z;
prt("best_z=%d,rate=%6.3f,games=%4d,playouts=%d,nodes=%d\n",
get81(best_z), pN->child[best_i].rate, max, all_playouts, node_num);
return best_z;
}
