/* Check if we can make a move along the fbook right away.
* Otherwise return pass. */
coord_t
fbook_check(struct board *board)
{
if (!board->fbook) return pass;
hash_t hi = board->hash;
coord_t cf = pass;
while (!is_pass(board->fbook->moves[hi & fbook_hash_mask])) {
if (board->fbook->hashes[hi & fbook_hash_mask] == board->hash) {
cf = board->fbook->moves[hi & fbook_hash_mask];
break;
}
hi++;
}
if (!is_pass(cf)) {
if (DEBUGL(1))
fprintf(stderr, "fbook match %"PRIhash":%"PRIhash"\n", board->hash, board->hash & fbook_hash_mask);
} else {
/* No match, also prevent further fbook usage
* until the next clear_board. */
if (DEBUGL(4))
fprintf(stderr, "fbook out %"PRIhash":%"PRIhash"\n", board->hash, board->hash & fbook_hash_mask);
fbook_done(board->fbook);
board->fbook = NULL;
}
return cf;
}
void
pattern_get(struct pattern_config *pc, struct pattern *p, struct board *b, struct move *m)
{
p->n = 0;
struct feature *f = &p->f[0];
/* TODO: We should match pretty much all of these features
* incrementally. */
/* FEAT_SPATIAL */
/* TODO */
assert(!pc->spat_max);
/* FEAT_PASS */
if (is_pass(m->coord)) {
f->id = FEAT_PASS; f->payload = 0;
f->payload |= (b->moves > 0 && is_pass(b->last_move.coord)) << PF_PASS_LASTPASS;
p->n++;
return;
}
/* FEAT_CAPTURE */
{
foreach_neighbor(b, m->coord, {
if (board_at(b, c) != stone_other(m->color))
continue;
group_t g = group_at(b, c);
if (!g || board_group_info(b, g).libs != 1)
continue;
/* Capture! */
f->id = FEAT_CAPTURE; f->payload = 0;
f->payload |= is_ladder(b, m->coord, g, true, true) << PF_CAPTURE_LADDER;
/* TODO: is_ladder() is too conservative in some
* very obvious situations, look at complete.gtp. */
/* TODO: PF_CAPTURE_RECAPTURE */
foreach_in_group(b, g) {
foreach_neighbor(b, c, {
assert(board_at(b, c) != S_NONE || c == m->coord);
if (board_at(b, c) != m->color)
continue;
group_t g = group_at(b, c);
if (!g || board_group_info(b, g).libs != 1)
continue;
/* A neighboring group of ours is in atari. */
f->payload |= 1 << PF_CAPTURE_ATARIDEF;
});
} foreach_in_group_end;
if (group_is_onestone(b, g)
&& neighbor_count_at(b, m->coord, stone_other(m->color))
+ neighbor_count_at(b, m->coord, S_OFFBOARD) == 4)
f->payload |= 1 << PF_CAPTURE_KO;
(f++, p->n++);
});
static coord_t *
replay_genmove(struct engine *e, struct board *b, struct time_info *ti, enum stone color, bool pass_all_alive)
{
struct replay *r = e->data;
struct playout_setup s; memset(&s, 0, sizeof(s));
coord_t coord = r->playout->choose(r->playout, &s, b, color);
if (!is_pass(coord)) {
struct move m;
m.coord = coord; m.color = color;
if (board_play(b, &m) >= 0)
goto have_move;
if (DEBUGL(2)) {
fprintf(stderr, "Pre-picked move %d,%d is ILLEGAL:\n",
coord_x(coord, b), coord_y(coord, b));
board_print(b, stderr);
}
}
/* Defer to uniformly random move choice. */
board_play_random(b, color, &coord, (ppr_permit) r->playout->permit, r->playout);
have_move:
if (!group_at(b, coord)) {
/* This was suicide. Just pass. */
/* XXX: We should check for non-suicide alternatives. */
return coord_pass();
}
return coord_copy(coord);
}
void
sgf_trace_semeai(const char *func, int str1, int str2, int move,
int result1, int result2, const char *message)
{
char buf[100];
sprintf(buf, "%s %c%d %c%d: ", func,
J(str1) + 'A' + (J(str1) >= 8), board_size - I(str1),
J(str2) + 'A' + (J(str2) >= 8), board_size - I(str2));
if (ON_BOARD(move))
sprintf(buf + strlen(buf), "%s %s %c%d",
result_to_string(result1), result_to_string(result2),
J(move) + 'A' + (J(move) >= 8), board_size - I(move));
else if (is_pass(move))
sprintf(buf + strlen(buf), "%s %s PASS",
result_to_string(result1), result_to_string(result2));
else
sprintf(buf + strlen(buf), "%s %s [%d]",
result_to_string(result1), result_to_string(result2),
move);
if (message)
sprintf(buf + strlen(buf), " (%s)", message);
sgftreeAddComment(sgf_dumptree, buf);
}
Move Move::mirror() const
{
if (is_pass())
return PASS;
int d = (direction() + (direction() & 1 ? 5 : 3)) & 7;
Rotation* rot = &block_set[block_id()]->rotations[d];
int new_x = y() + rot->offset_x;
int new_y = x() + rot->offset_y;
return Move(new_x, new_y, rot->piece->id);
}
const char* Move::fourcc() const
{
static char buf[5];
if (is_pass())
strcpy(buf, "----");
else
sprintf(buf, "%2X%c%d",
(m_ & 0xff) + 0x11, 'u' - block_id(), direction());
return buf;
}
static int
computer_move(Gameinfo *gameinfo, int *passes)
{
int move;
float move_value;
int resign;
int resignation_declined = 0;
float upper_bound, lower_bound;
init_sgf(gameinfo);
/* Generate computer move. */
if (autolevel_on)
adjust_level_offset(gameinfo->to_move);
move = genmove(gameinfo->to_move, &move_value, &resign);
if (resignation_allowed && resign) {
int state = ascii_endgame(gameinfo, 2);
if (state != -1)
return state;
/* The opponent declined resignation. Remember not to resign again. */
resignation_allowed = 0;
resignation_declined = 1;
}
if (showscore) {
gnugo_estimate_score(&upper_bound, &lower_bound);
current_score_estimate = (int) ((lower_bound + upper_bound) / 2.0);
}
mprintf("%s(%d): %1m\n", color_to_string(gameinfo->to_move),
movenum + 1, move);
if (is_pass(move))
(*passes)++;
else
*passes = 0;
gnugo_play_move(move, gameinfo->to_move);
sgffile_add_debuginfo(sgftree.lastnode, move_value);
sgftreeAddPlay(&sgftree, gameinfo->to_move, I(move), J(move));
if (resignation_declined)
sgftreeAddComment(&sgftree, "GNU Go resignation was declined");
sgffile_output(&sgftree);
gameinfo->to_move = OTHER_COLOR(gameinfo->to_move);
return 0;
}
char *
coord2sstr(coord_t c, struct board *board)
{
static char *b;
static char bl[10][4];
static int bi;
if (is_pass(c)) {
return "pass";
} else if (is_resign(c)) {
return "resign";
} else {
/* Some GTP servers are broken and won't grok lowercase coords */
b = bl[bi]; bi = (bi + 1) % 10;
snprintf(b, 4, "%c%d", toupper(asdf[coord_x(c, board) - 1]), coord_y(c, board));
return b;
}
}
static floating_t
pattern_rate_move(pattern_config_t *pc,
board_t *b, move_t *m,
pattern_t *pat, ownermap_t *ownermap, bool locally)
{
floating_t prob = NAN;
if (is_pass(m->coord)) return prob;
if (!board_is_valid_play_no_suicide(b, m->color, m->coord)) return prob;
pattern_match(pc, pat, b, m, ownermap, locally);
prob = pattern_gamma(pc, pat);
//if (DEBUGL(5)) {
// char buf[256]; pattern2str(buf, pat);
// fprintf(stderr, "=> move %s pattern %s prob %.3f\n", coord2sstr(m->coord), buf, prob);
//}
return prob;
}
void
load_and_analyze_sgf_file(Gameinfo *gameinfo)
{
SGFTree sgftree;
int i, j;
int next;
int move_val;
next = gameinfo->to_move;
sgftree = gameinfo->game_record;
move_val = gnugo_genmove(&i, &j, next);
if (is_pass(POS(i, j)))
gprintf("%s move: PASS!\n", next == WHITE ? "white (O)" : "black (X)");
else
gprintf("%s move %m\n", next == WHITE ? "white (O)" : "black (X)", i, j);
gnugo_play_move(i, j, next);
sgftreeAddPlay(&sgftree, next, i, j);
sgftreeAddComment(&sgftree, "load and analyze mode");
sgffile_add_debuginfo(sgftree.lastnode, move_val);
sgffile_output(&sgftree);
}
/* --------------------------------------------------------------*/
void
play_gmp(Gameinfo *gameinfo, int simplified)
{
SGFTree sgftree;
Gmp *ge;
GmpResult message;
const char *error;
int i, j;
int passes = 0; /* two passes and its over */
int to_move; /* who's turn is next ? */
int mycolor = -1; /* who has which color */
int yourcolor;
if (gameinfo->computer_player == WHITE)
mycolor = 1;
else if (gameinfo->computer_player == BLACK)
mycolor = 0;
sgftree_clear(&sgftree);
sgftreeCreateHeaderNode(&sgftree, board_size, komi, gameinfo->handicap);
ge = gmp_create(0, 1);
TRACE("board size=%d\n", board_size);
/*
* The specification of the go modem protocol doesn't even discuss
* komi. So we have to guess the komi. If the komi is set on the
* command line, keep it. Otherwise, its value will be 0.0 and we
* use 5.5 in an even game, 0.5 otherwise.
*/
if (komi == 0.0) {
if (gameinfo->handicap == 0)
komi = 5.5;
else
komi = 0.5;
}
if (!simplified) {
/* Leave all the -1's so the client can negotiate the game parameters. */
if (chinese_rules)
gmp_startGame(ge, -1, -1, 5.5, -1, mycolor, 0);
else
gmp_startGame(ge, -1, -1, 5.5, 0, mycolor, 0);
}
else {
gmp_startGame(ge, board_size, gameinfo->handicap,
komi, chinese_rules, mycolor, 1);
}
do {
message = gmp_check(ge, 1, NULL, NULL, &error);
} while (message == gmp_nothing || message == gmp_reset);
if (message == gmp_err) {
fprintf(stderr, "gnugo-gmp: Error \"%s\" occurred.\n", error);
exit(EXIT_FAILURE);
}
else if (message != gmp_newGame) {
fprintf(stderr, "gnugo-gmp: Expecting a newGame, got %s\n",
gmp_resultString(message));
exit(EXIT_FAILURE);
}
gameinfo->handicap = gmp_handicap(ge);
if (!check_boardsize(gmp_size(ge), stderr))
exit(EXIT_FAILURE);
gnugo_clear_board(gmp_size(ge));
/* Let's pretend GMP knows about komi in case something will ever change. */
komi = gmp_komi(ge);
#if ORACLE
if (metamachine && oracle_exists)
oracle_clear_board(board_size);
#endif
sgfOverwritePropertyInt(sgftree.root, "SZ", board_size);
TRACE("size=%d, handicap=%d, komi=%f\n", board_size,
gameinfo->handicap, komi);
if (gameinfo->handicap)
to_move = WHITE;
else
to_move = BLACK;
if (gmp_iAmWhite(ge)) {
mycolor = WHITE; /* computer white */
yourcolor = BLACK; /* human black */
}
else {
mycolor = BLACK;
yourcolor = WHITE;
}
gameinfo->computer_player = mycolor;
sgf_write_header(sgftree.root, 1, get_random_seed(), komi,
gameinfo->handicap, get_level(), chinese_rules);
gameinfo->handicap = gnugo_sethand(gameinfo->handicap, sgftree.root);
sgfOverwritePropertyInt(sgftree.root, "HA", gameinfo->handicap);
/* main GMP loop */
while (passes < 2) {
if (to_move == yourcolor) {
int move;
/* Get opponent's move from gmp client. */
message = gmp_check(ge, 1, &j, &i, &error);
if (message == gmp_err) {
fprintf(stderr, "GNU Go: Sorry, error from gmp client\n");
sgftreeAddComment(&sgftree, "got error from gmp client");
sgffile_output(&sgftree);
return;
}
if (message == gmp_undo) {
int k;
assert(j > 0);
for (k = 0; k < j; k++) {
if (!undo_move(1)) {
fprintf(stderr, "GNU Go: play_gmp UNDO: can't undo %d moves\n",
j - k);
break;
}
sgftreeAddComment(&sgftree, "undone");
sgftreeBack(&sgftree);
to_move = OTHER_COLOR(to_move);
}
continue;
}
if (message == gmp_pass) {
passes++;
move = PASS_MOVE;
}
else {
passes = 0;
move = POS(i, j);
}
TRACE("\nyour move: %1m\n\n", move);
sgftreeAddPlay(&sgftree, to_move, I(move), J(move));
gnugo_play_move(move, yourcolor);
sgffile_output(&sgftree);
}
else {
/* Generate my next move. */
float move_value;
int move;
if (autolevel_on)
adjust_level_offset(mycolor);
move = genmove(mycolor, &move_value, NULL);
gnugo_play_move(move, mycolor);
sgffile_add_debuginfo(sgftree.lastnode, move_value);
if (is_pass(move)) {
/* pass */
sgftreeAddPlay(&sgftree, to_move, -1, -1);
gmp_sendPass(ge);
++passes;
}
else {
/* not pass */
sgftreeAddPlay(&sgftree, to_move, I(move), J(move));
gmp_sendMove(ge, J(move), I(move));
passes = 0;
TRACE("\nmy move: %1m\n\n", move);
}
sgffile_add_debuginfo(sgftree.lastnode, 0.0);
sgffile_output(&sgftree);
}
to_move = OTHER_COLOR(to_move);
}
/* two passes: game over */
gmp_sendPass(ge);
if (!quiet)
fprintf(stderr, "Game over - waiting for client to shut us down\n");
who_wins(mycolor, stderr);
if (showtime) {
gprintf("\nSLOWEST MOVE: %d at %1m ", slowest_movenum, slowest_move);
fprintf(stderr, "(%.2f seconds)\n", slowest_time);
fprintf(stderr, "\nAVERAGE TIME: %.2f seconds per move\n",
total_time / movenum);
fprintf(stderr, "\nTOTAL TIME: %.2f seconds\n",
total_time);
}
/* play_gmp() does not return to main(), therefore the score
* writing code is here.
*/
{
float score = gnugo_estimate_score(NULL, NULL);
sgfWriteResult(sgftree.root, score, 1);
}
sgffile_output(&sgftree);
if (!simplified) {
/* We hang around here until cgoban asks us to go, since
* sometimes cgoban crashes if we exit first.
*
* FIXME: Check if this is still needed. I made it dependand on
* `simplifed' just to avoid changes in GMP mode.
*/
while (1) {
message = gmp_check(ge, 1, &j, &i, &error);
if (!quiet)
fprintf(stderr, "Message %d from gmp\n", message);
if (message == gmp_err)
break;
}
}
#if ORACLE
if (metamachine && oracle_exists)
dismiss_oracle();
#endif
if (!quiet)
fprintf(stderr, "gnugo going down\n");
}
int
uct_playout(struct uct *u, struct board *b, enum stone player_color, struct tree *t)
{
struct board b2;
board_copy(&b2, b);
struct playout_amafmap amaf;
amaf.gamelen = amaf.game_baselen = 0;
/* Walk the tree until we find a leaf, then expand it and do
* a random playout. */
struct tree_node *n = t->root;
enum stone node_color = stone_other(player_color);
assert(node_color == t->root_color);
/* Make sure the root node is expanded. */
if (tree_leaf_node(n) && !__sync_lock_test_and_set(&n->is_expanded, 1))
tree_expand_node(t, n, &b2, player_color, u, 1);
/* Tree descent history. */
/* XXX: This is somewhat messy since @n and descent[dlen-1].node are
* redundant. */
struct uct_descent descent[DESCENT_DLEN];
descent[0].node = n; descent[0].lnode = NULL;
int dlen = 1;
/* Total value of the sequence. */
struct move_stats seq_value = { .playouts = 0 };
/* The last "significant" node along the descent (i.e. node
* with higher than configured number of playouts). For black
* and white. */
struct tree_node *significant[2] = { NULL, NULL };
if (n->u.playouts >= u->significant_threshold)
significant[node_color - 1] = n;
int result;
int pass_limit = (board_size(&b2) - 2) * (board_size(&b2) - 2) / 2;
int passes = is_pass(b->last_move.coord) && b->moves > 0;
/* debug */
static char spaces[] = "\0 ";
/* /debug */
if (UDEBUGL(8))
fprintf(stderr, "--- UCT walk with color %d\n", player_color);
while (!tree_leaf_node(n) && passes < 2) {
spaces[dlen - 1] = ' '; spaces[dlen] = 0;
/*** Choose a node to descend to: */
/* Parity is chosen already according to the child color, since
* it is applied to children. */
node_color = stone_other(node_color);
int parity = (node_color == player_color ? 1 : -1);
assert(dlen < DESCENT_DLEN);
descent[dlen] = descent[dlen - 1];
if (u->local_tree && (!descent[dlen].lnode || descent[dlen].node->d >= u->tenuki_d)) {
/* Start new local sequence. */
/* Remember that node_color already holds color of the
* to-be-found child. */
descent[dlen].lnode = node_color == S_BLACK ? t->ltree_black : t->ltree_white;
}
if (!u->random_policy_chance || fast_random(u->random_policy_chance))
u->policy->descend(u->policy, t, &descent[dlen], parity, b2.moves > pass_limit);
else
u->random_policy->descend(u->random_policy, t, &descent[dlen], parity, b2.moves > pass_limit);
/*** Perform the descent: */
if (descent[dlen].node->u.playouts >= u->significant_threshold) {
significant[node_color - 1] = descent[dlen].node;
}
seq_value.playouts += descent[dlen].value.playouts;
seq_value.value += descent[dlen].value.value * descent[dlen].value.playouts;
n = descent[dlen++].node;
assert(n == t->root || n->parent);
if (UDEBUGL(7))
fprintf(stderr, "%s+-- UCT sent us to [%s:%d] %d,%f\n",
spaces, coord2sstr(node_coord(n), t->board),
node_coord(n), n->u.playouts,
tree_node_get_value(t, parity, n->u.value));
/* Add virtual loss if we need to; this is used to discourage
* other threads from visiting this node in case of multiple
* threads doing the tree search. */
if (u->virtual_loss)
stats_add_result(&n->u, node_color == S_BLACK ? 0.0 : 1.0, u->virtual_loss);
assert(node_coord(n) >= -1);
record_amaf_move(&amaf, node_coord(n));
struct move m = { node_coord(n), node_color };
int res = board_play(&b2, &m);
if (res < 0 || (!is_pass(m.coord) && !group_at(&b2, m.coord)) /* suicide */
|| b2.superko_violation) {
if (UDEBUGL(4)) {
for (struct tree_node *ni = n; ni; ni = ni->parent)
fprintf(stderr, "%s<%"PRIhash"> ", coord2sstr(node_coord(ni), t->board), ni->hash);
fprintf(stderr, "marking invalid %s node %d,%d res %d group %d spk %d\n",
stone2str(node_color), coord_x(node_coord(n),b), coord_y(node_coord(n),b),
res, group_at(&b2, m.coord), b2.superko_violation);
}
n->hints |= TREE_HINT_INVALID;
result = 0;
goto end;
}
if (is_pass(node_coord(n)))
passes++;
else
passes = 0;
enum stone next_color = stone_other(node_color);
/* We need to make sure only one thread expands the node. If
* we are unlucky enough for two threads to meet in the same
* node, the latter one will simply do another simulation from
* the node itself, no big deal. t->nodes_size may exceed
* the maximum in multi-threaded case but not by much so it's ok.
* The size test must be before the test&set not after, to allow
* expansion of the node later if enough nodes have been freed. */
if (tree_leaf_node(n)
&& n->u.playouts - u->virtual_loss >= u->expand_p && t->nodes_size < u->max_tree_size
&& !__sync_lock_test_and_set(&n->is_expanded, 1))
tree_expand_node(t, n, &b2, next_color, u, -parity);
}
amaf.game_baselen = amaf.gamelen;
if (t->use_extra_komi && u->dynkomi->persim) {
b2.komi += round(u->dynkomi->persim(u->dynkomi, &b2, t, n));
}
if (passes >= 2) {
/* XXX: No dead groups support. */
floating_t score = board_official_score(&b2, NULL);
/* Result from black's perspective (no matter who
* the player; black's perspective is always
* what the tree stores. */
result = - (score * 2);
if (UDEBUGL(5))
fprintf(stderr, "[%d..%d] %s p-p scoring playout result %d (W %f)\n",
player_color, node_color, coord2sstr(node_coord(n), t->board), result, score);
if (UDEBUGL(6))
board_print(&b2, stderr);
board_ownermap_fill(&u->ownermap, &b2);
} else { // assert(tree_leaf_node(n));
/* In case of parallel tree search, the assertion might
* not hold if two threads chew on the same node. */
result = uct_leaf_node(u, &b2, player_color, &amaf, descent, &dlen, significant, t, n, node_color, spaces);
}
if (u->policy->wants_amaf && u->playout_amaf_cutoff) {
unsigned int cutoff = amaf.game_baselen;
cutoff += (amaf.gamelen - amaf.game_baselen) * u->playout_amaf_cutoff / 100;
amaf.gamelen = cutoff;
}
/* Record the result. */
assert(n == t->root || n->parent);
floating_t rval = scale_value(u, b, result);
u->policy->update(u->policy, t, n, node_color, player_color, &amaf, &b2, rval);
if (t->use_extra_komi) {
stats_add_result(&u->dynkomi->score, result / 2, 1);
stats_add_result(&u->dynkomi->value, rval, 1);
}
if (u->local_tree && n->parent && !is_pass(node_coord(n)) && dlen > 0) {
/* Get the local sequences and record them in ltree. */
/* We will look for sequence starts in our descent
* history, then run record_local_sequence() for each
* found sequence start; record_local_sequence() may
* pick longer sequences from descent history then,
* which is expected as it will create new lnodes. */
enum stone seq_color = player_color;
/* First move always starts a sequence. */
record_local_sequence(u, t, &b2, descent, dlen, 1, seq_color);
seq_color = stone_other(seq_color);
for (int dseqi = 2; dseqi < dlen; dseqi++, seq_color = stone_other(seq_color)) {
if (u->local_tree_allseq) {
/* We are configured to record all subsequences. */
record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
continue;
}
if (descent[dseqi].node->d >= u->tenuki_d) {
/* Tenuki! Record the fresh sequence. */
record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
continue;
}
if (descent[dseqi].lnode && !descent[dseqi].lnode) {
/* Record result for in-descent picked sequence. */
record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
continue;
}
}
}
end:
/* We need to undo the virtual loss we added during descend. */
if (u->virtual_loss) {
floating_t loss = node_color == S_BLACK ? 0.0 : 1.0;
for (; n->parent; n = n->parent) {
stats_rm_result(&n->u, loss, u->virtual_loss);
loss = 1.0 - loss;
}
}
board_done_noalloc(&b2);
return result;
}
int
uct_playouts(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti)
{
int i;
if (ti && ti->dim == TD_GAMES) {
for (i = 0; t->root->u.playouts <= ti->len.games && !uct_halt; i++)
uct_playout(u, b, color, t);
} else {
for (i = 0; !uct_halt; i++)
uct_playout(u, b, color, t);
}
return i;
}
static void
record_local_sequence(struct uct *u, struct tree *t, struct board *endb,
struct uct_descent *descent, int dlen, int di,
enum stone seq_color)
{
#define LTREE_DEBUG if (UDEBUGL(6))
/* Ignore pass sequences. */
if (is_pass(node_coord(descent[di].node)))
return;
LTREE_DEBUG board_print(endb, stderr);
LTREE_DEBUG fprintf(stderr, "recording local %s sequence: ",
stone2str(seq_color));
/* Sequences starting deeper are less relevant in general. */
int pval = LTREE_PLAYOUTS_MULTIPLIER;
if (u->local_tree && u->local_tree_depth_decay > 0)
pval = ((floating_t) pval) / pow(u->local_tree_depth_decay, di - 1);
if (!pval) {
LTREE_DEBUG fprintf(stderr, "too deep @%d\n", di);
return;
}
/* Pick the right local tree root... */
struct tree_node *lnode = seq_color == S_BLACK ? t->ltree_black : t->ltree_white;
lnode->u.playouts++;
double sval = 0.5;
if (u->local_tree_rootgoal) {
sval = local_value(u, endb, node_coord(descent[di].node), seq_color);
LTREE_DEBUG fprintf(stderr, "(goal %s[%s %1.3f][%d]) ",
coord2sstr(node_coord(descent[di].node), t->board),
stone2str(seq_color), sval, descent[di].node->d);
}
/* ...and record the sequence. */
int di0 = di;
while (di < dlen && (di == di0 || descent[di].node->d < u->tenuki_d)) {
enum stone color = (di - di0) % 2 ? stone_other(seq_color) : seq_color;
double rval;
if (u->local_tree_rootgoal)
rval = sval;
else
rval = local_value(u, endb, node_coord(descent[di].node), color);
LTREE_DEBUG fprintf(stderr, "%s[%s %1.3f][%d] ",
coord2sstr(node_coord(descent[di].node), t->board),
stone2str(color), rval, descent[di].node->d);
lnode = tree_get_node(t, lnode, node_coord(descent[di++].node), true);
assert(lnode);
stats_add_result(&lnode->u, rval, pval);
}
/* Add lnode for tenuki (pass) if we descended further. */
if (di < dlen) {
double rval = u->local_tree_rootgoal ? sval : 0.5;
LTREE_DEBUG fprintf(stderr, "pass ");
lnode = tree_get_node(t, lnode, pass, true);
assert(lnode);
stats_add_result(&lnode->u, rval, pval);
}
LTREE_DEBUG fprintf(stderr, "\n");
}
static coord_t *
montecarlo_genmove(struct engine *e, struct board *b, struct time_info *ti, enum stone color, bool pass_all_alive)
{
struct montecarlo *mc = e->data;
if (ti->dim == TD_WALLTIME) {
fprintf(stderr, "Warning: TD_WALLTIME time mode not supported, resetting to defaults.\n");
ti->period = TT_NULL;
}
if (ti->period == TT_NULL) {
ti->period = TT_MOVE;
ti->dim = TD_GAMES;
ti->len.games = MC_GAMES;
}
struct time_stop stop;
time_stop_conditions(ti, b, 20, 40, 3.0, &stop);
/* resign when the hope for win vanishes */
coord_t top_coord = resign;
floating_t top_ratio = mc->resign_ratio;
/* We use [0] for pass. Normally, this is an inaccessible corner
* of board margin. */
struct move_stat moves[board_size2(b)];
memset(moves, 0, sizeof(moves));
int losses = 0;
int i, superko = 0, good_games = 0;
for (i = 0; i < stop.desired.playouts; i++) {
assert(!b->superko_violation);
struct board b2;
board_copy(&b2, b);
coord_t coord;
board_play_random(&b2, color, &coord, NULL, NULL);
if (!is_pass(coord) && !group_at(&b2, coord)) {
/* Multi-stone suicide. We play chinese rules,
* so we can't consider this. (Note that we
* unfortunately still consider this in playouts.) */
if (DEBUGL(4)) {
fprintf(stderr, "SUICIDE DETECTED at %d,%d:\n", coord_x(coord, b), coord_y(coord, b));
board_print(b, stderr);
}
continue;
}
if (DEBUGL(3))
fprintf(stderr, "[%d,%d color %d] playing random game\n", coord_x(coord, b), coord_y(coord, b), color);
struct playout_setup ps = { .gamelen = mc->gamelen };
int result = play_random_game(&ps, &b2, color, NULL, NULL, mc->playout);
board_done_noalloc(&b2);
if (result == 0) {
/* Superko. We just ignore this playout.
* And play again. */
if (unlikely(superko > 2 * stop.desired.playouts)) {
/* Uhh. Triple ko, or something? */
if (MCDEBUGL(0))
fprintf(stderr, "SUPERKO LOOP. I will pass. Did we hit triple ko?\n");
goto pass_wins;
}
/* This playout didn't count; we should not
* disadvantage moves that lead to a superko.
* And it is supposed to be rare. */
i--, superko++;
continue;
}
if (MCDEBUGL(3))
fprintf(stderr, "\tresult for other player: %d\n", result);
int pos = is_pass(coord) ? 0 : coord;
good_games++;
moves[pos].games++;
losses += result > 0;
moves[pos].wins += 1 - (result > 0);
if (unlikely(!losses && i == mc->loss_threshold)) {
/* We played out many games and didn't lose once yet.
* This game is over. */
break;
}
}
if (!good_games) {
/* No moves to try??? */
if (MCDEBUGL(0)) {
fprintf(stderr, "OUT OF MOVES! I will pass. But how did this happen?\n");
board_print(b, stderr);
}
pass_wins:
top_coord = pass; top_ratio = 0.5;
goto move_found;
}
foreach_point(b) {
if (b->moves < 3) {
/* Simple heuristic: avoid opening too low. Do not
* play on second or first line as first white or
* first two black moves.*/
if (coord_x(c, b) < 3 || coord_x(c, b) > board_size(b) - 4
|| coord_y(c, b) < 3 || coord_y(c, b) > board_size(b) - 4)
continue;
}
floating_t ratio = (floating_t) moves[c].wins / moves[c].games;
/* Since pass is [0,0], we will pass only when we have nothing
* better to do. */
if (ratio >= top_ratio) {
top_ratio = ratio;
top_coord = c == 0 ? pass : c;
}
} foreach_point_end;
if (MCDEBUGL(2)) {
board_stats_print(b, moves, stderr);
}
move_found:
if (MCDEBUGL(1))
fprintf(stderr, "*** WINNER is %d,%d with score %1.4f (%d games, %d superko)\n", coord_x(top_coord, b), coord_y(top_coord, b), top_ratio, i, superko);
return coord_copy(top_coord);
}
static void
montecarlo_done(struct engine *e)
{
struct montecarlo *mc = e->data;
playout_policy_done(mc->playout);
joseki_done(mc->jdict);
}
struct montecarlo *
montecarlo_state_init(char *arg, struct board *b)
{
struct montecarlo *mc = calloc2(1, sizeof(struct montecarlo));
mc->debug_level = 1;
mc->gamelen = MC_GAMELEN;
mc->jdict = joseki_load(b->size);
if (arg) {
char *optspec, *next = arg;
while (*next) {
optspec = next;
next += strcspn(next, ",");
if (*next) { *next++ = 0; } else { *next = 0; }
char *optname = optspec;
char *optval = strchr(optspec, '=');
if (optval) *optval++ = 0;
if (!strcasecmp(optname, "debug")) {
if (optval)
mc->debug_level = atoi(optval);
else
mc->debug_level++;
} else if (!strcasecmp(optname, "gamelen") && optval) {
mc->gamelen = atoi(optval);
} else if (!strcasecmp(optname, "playout") && optval) {
char *playoutarg = strchr(optval, ':');
if (playoutarg)
*playoutarg++ = 0;
if (!strcasecmp(optval, "moggy")) {
mc->playout = playout_moggy_init(playoutarg, b, mc->jdict);
} else if (!strcasecmp(optval, "light")) {
mc->playout = playout_light_init(playoutarg, b);
} else {
fprintf(stderr, "MonteCarlo: Invalid playout policy %s\n", optval);
}
} else {
fprintf(stderr, "MonteCarlo: Invalid engine argument %s or missing value\n", optname);
}
}
}
if (!mc->playout)
mc->playout = playout_light_init(NULL, b);
mc->playout->debug_level = mc->debug_level;
mc->resign_ratio = 0.1; /* Resign when most games are lost. */
mc->loss_threshold = 5000; /* Stop reading if no loss encountered in first 5000 games. */
return mc;
}
struct engine *
engine_montecarlo_init(char *arg, struct board *b)
{
struct montecarlo *mc = montecarlo_state_init(arg, b);
struct engine *e = calloc2(1, sizeof(struct engine));
e->name = "MonteCarlo";
e->comment = "I'm playing in Monte Carlo. When we both pass, I will consider all the stones on the board alive. If you are reading this, write 'yes'. Please bear with me at the game end, I need to fill the whole board; if you help me, we will both be happier. Filling the board will not lose points (NZ rules).";
e->genmove = montecarlo_genmove;
e->done = montecarlo_done;
e->data = mc;
return e;
}
struct fbook *
fbook_init(char *filename, struct board *b)
{
if (fbcache && fbcache->bsize == board_size(b)
&& fbcache->handicap == b->handicap)
return fbcache;
FILE *f = fopen(filename, "r");
if (!f) {
perror(filename);
return NULL;
}
struct fbook *fbook = calloc(1, sizeof(*fbook));
fbook->bsize = board_size(b);
fbook->handicap = b->handicap;
/* We do not set handicap=1 in case of too low komi on purpose;
* we want to go with the no-handicap fbook for now. */
for (int i = 0; i < 1<<fbook_hash_bits; i++)
fbook->moves[i] = pass;
if (DEBUGL(1))
fprintf(stderr, "Loading opening fbook %s...\n", filename);
/* Scratch board where we lay out the sequence;
* one for each transposition. */
struct board *bs[8];
for (int i = 0; i < 8; i++) {
bs[i] = board_init(NULL);
board_resize(bs[i], fbook->bsize - 2);
}
char linebuf[1024];
while (fgets(linebuf, sizeof(linebuf), f)) {
char *line = linebuf;
linebuf[strlen(linebuf) - 1] = 0; // chop
/* Format of line is:
* BSIZE COORD COORD COORD... | COORD
* BSIZE/HANDI COORD COORD COORD... | COORD
* We descend up to |, then add the new node
* with value minimax(1000), forcing UCT to
* always pick that node immediately. */
int bsize = strtol(line, &line, 10);
if (bsize != fbook->bsize - 2)
continue;
int handi = 0;
if (*line == '/') {
line++;
handi = strtol(line, &line, 10);
}
if (handi != fbook->handicap)
continue;
while (isspace(*line)) line++;
for (int i = 0; i < 8; i++) {
board_clear(bs[i]);
bs[i]->last_move.color = S_WHITE;
}
while (*line != '|') {
coord_t *c = str2coord(line, fbook->bsize);
for (int i = 0; i < 8; i++) {
coord_t coord = coord_transform(b, *c, i);
struct move m = { .coord = coord, .color = stone_other(bs[i]->last_move.color) };
int ret = board_play(bs[i], &m);
assert(ret >= 0);
}
coord_done(c);
while (!isspace(*line)) line++;
while (isspace(*line)) line++;
}
line++;
while (isspace(*line)) line++;
/* In case of multiple candidates, pick one with
* exponentially decreasing likelihood. */
while (strchr(line, ' ') && fast_random(2)) {
line = strchr(line, ' ');
while (isspace(*line)) line++;
// fprintf(stderr, "<%s> skip to %s\n", linebuf, line);
}
coord_t *c = str2coord(line, fbook->bsize);
for (int i = 0; i < 8; i++) {
coord_t coord = coord_transform(b, *c, i);
#if 0
char conflict = is_pass(fbook->moves[bs[i]->hash & fbook_hash_mask]) ? '+' : 'C';
if (conflict == 'C')
for (int j = 0; j < i; j++)
if (bs[i]->hash == bs[j]->hash)
conflict = '+';
if (conflict == 'C') {
hash_t hi = bs[i]->hash;
while (!is_pass(fbook->moves[hi & fbook_hash_mask]) && fbook->hashes[hi & fbook_hash_mask] != bs[i]->hash)
hi++;
if (fbook->hashes[hi & fbook_hash_mask] == bs[i]->hash)
hi = 'c';
}
fprintf(stderr, "%c %"PRIhash":%"PRIhash" (<%d> %s)\n", conflict,
bs[i]->hash & fbook_hash_mask, bs[i]->hash, i, linebuf);
#endif
hash_t hi = bs[i]->hash;
while (!is_pass(fbook->moves[hi & fbook_hash_mask]) && fbook->hashes[hi & fbook_hash_mask] != bs[i]->hash)
hi++;
fbook->moves[hi & fbook_hash_mask] = coord;
fbook->hashes[hi & fbook_hash_mask] = bs[i]->hash;
fbook->movecnt++;
}
coord_done(c);
}
for (int i = 0; i < 8; i++) {
board_done(bs[i]);
}
fclose(f);
if (!fbook->movecnt) {
/* Empty book is not worth the hassle. */
fbook_done(fbook);
return NULL;
}
struct fbook *fbold = fbcache;
fbcache = fbook;
if (fbold)
fbook_done(fbold);
return fbook;
}
void fbook_done(struct fbook *fbook)
{
if (fbook != fbcache)
free(fbook);
}
