static coord_t * random_genmove(struct engine *e, struct board *b, struct time_info *ti, enum stone color, bool pass_all_alive) { /* Play a random coordinate. However, we must also guard * against suicide moves; repeat playing while it's a suicide * unless we keep suiciding; in that case, we probably don't * have any other moves available and we pass. */ coord_t coord; int i = 0; bool suicide = false; do { /* board_play_random() actually plays the move too; * this is desirable for MC simulations but not within * the genmove. Make a scratch new board for it. */ struct board b2; board_copy(&b2, b); board_play_random(&b2, color, &coord, NULL, NULL); suicide = (coord != pass && !group_at(&b2, coord)); board_done_noalloc(&b2); } while (suicide && i++ < 100); return coord_copy(suicide ? pass : coord); }
void board_done(struct board *board) { board_done_noalloc(board); free(board); }
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 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; }
/* Syntax: * moggy status (last_move) coord [coord...] * Play number of random games starting from last_move * * moggy status coord [coord...] * moggy status (b) coord [coord...] * Black to play, pick random white last move * * moggy status (w) coord [coord...] * White to play, pick random black last move */ static bool test_moggy_status(struct board *board, char *arg) { int games = 4000; coord_t status_at[10]; int n = 0; enum stone color = S_BLACK; int pick_random = true; // Pick random last move for each game while (*arg && *arg != '#') { if (*arg == ' ' || *arg == '\t') { arg++; continue; } if (!strncmp(arg, "(b)", 3)) color = S_BLACK; else if (!strncmp(arg, "(w)", 3)) color = S_WHITE; else if (*arg == '(') { /* Optional "(last_move)" argument */ arg++; assert(isalpha(*arg)); pick_random = false; struct move last; last.coord = str2scoord(arg, board_size(board)); last.color = board_at(board, last.coord); assert(last.color == S_BLACK || last.color == S_WHITE); color = stone_other(last.color); board->last_move = last; } else { assert(isalpha(*arg)); status_at[n++] = str2scoord(arg, board_size(board)); } arg += strcspn(arg, " \t"); } board_print(board, stderr); if (DEBUGL(1)) { printf("moggy status "); for (int i = 0; i < n; i++) printf("%s%s", coord2sstr(status_at[i], board), (i != n-1 ? " " : "")); printf(", %s to play. Playing %i games %s...\n", stone2str(color), games, (pick_random ? "(random last move) " : "")); } struct playout_policy *policy = playout_moggy_init(NULL, board, NULL); struct playout_setup setup = { .gamelen = MAX_GAMELEN }; struct board_ownermap ownermap; ownermap.playouts = 0; ownermap.map = malloc2(board_size2(board) * sizeof(ownermap.map[0])); memset(ownermap.map, 0, board_size2(board) * sizeof(ownermap.map[0])); /* Get final status estimate after a number of moggy games */ int wr = 0; double time_start = time_now(); for (int i = 0; i < games; i++) { struct board b; board_copy(&b, board); if (pick_random) pick_random_last_move(&b, color); int score = play_random_game(&setup, &b, color, NULL, &ownermap, policy); if (color == S_WHITE) score = -score; wr += (score > 0); board_done_noalloc(&b); } double elapsed = time_now() - time_start; printf("moggy status in %.1fs, %i games/s\n\n", elapsed, (int)((float)games / elapsed)); int wr_black = wr * 100 / games; int wr_white = (games - wr) * 100 / games; if (wr_black > wr_white) printf("Winrate: [ black %i%% ] white %i%%\n\n", wr_black, wr_white); else printf("Winrate: black %i%% [ white %i%% ]\n\n", wr_black, wr_white); board_print_ownermap(board, stderr, &ownermap); for (int i = 0; i < n; i++) { coord_t c = status_at[i]; enum stone color = (ownermap.map[c][S_BLACK] > ownermap.map[c][S_WHITE] ? S_BLACK : S_WHITE); fprintf(stderr, "%3s owned by %s: %i%%\n", coord2sstr(c, board), stone2str(color), ownermap.map[c][color] * 100 / ownermap.playouts); } free(ownermap.map); playout_policy_done(policy); return true; // Not much of a unit test right now =) }