int main() {
char buf[100];
// After that we use moggy to analyze it. (With default parameters).
void* policy = playout_moggy_init(NULL);
// Put a few random moves and ask mogo to play it out.
AllMoves all_moves;
GroupId4 ids;
const int num_round = 10000;
Board b;
timeit
for (int j = 0; j < num_round; j++) {
printf("Round = %d/%d\n", j, num_round);
ClearBoard(&b);
//for (int i = 0; i < rand() % 100 + 1; ++i) {
for (int i = 0; i < 0; ++i) {
FindAllCandidateMoves(&b, b._next_player, 0, &all_moves);
// Randomly find one move.
int idx = rand() % all_moves.num_moves;
if (TryPlay2(&b, all_moves.moves[idx], &ids)) {
Play(&b, &ids);
}
}
ShowBoard(&b, SHOW_LAST_MOVE);
play_random_game(policy, NULL, NULL, &b, -1, FALSE);
ShowBoard(&b, SHOW_LAST_MOVE);
}
endtime
playout_moggy_destroy(policy);
return 0;
}
static int
uct_leaf_node(struct uct *u, struct board *b, enum stone player_color,
struct playout_amafmap *amaf,
struct uct_descent *descent, int *dlen,
struct tree_node *significant[2],
struct tree *t, struct tree_node *n, enum stone node_color,
char *spaces)
{
enum stone next_color = stone_other(node_color);
int parity = (next_color == player_color ? 1 : -1);
if (UDEBUGL(7))
fprintf(stderr, "%s*-- UCT playout #%d start [%s] %f\n",
spaces, n->u.playouts, coord2sstr(node_coord(n), t->board),
tree_node_get_value(t, parity, n->u.value));
struct uct_playout_callback upc = {
.uct = u,
.tree = t,
/* TODO: Don't necessarily restart the sequence walk when
* entering playout. */
.lnode = NULL,
};
struct playout_setup ps = {
.gamelen = u->gamelen,
.mercymin = u->mercymin,
.prepolicy_hook = uct_playout_prepolicy,
.postpolicy_hook = uct_playout_postpolicy,
.hook_data = &upc,
};
int result = play_random_game(&ps, b, next_color,
u->playout_amaf ? amaf : NULL,
&u->ownermap, u->playout);
if (next_color == S_WHITE) {
/* We need the result from black's perspective. */
result = - result;
}
if (UDEBUGL(7))
fprintf(stderr, "%s -- [%d..%d] %s random playout result %d\n",
spaces, player_color, next_color, coord2sstr(node_coord(n), t->board), result);
return result;
}
static floating_t
scale_value(struct uct *u, struct board *b, int result)
{
floating_t rval = result > 0 ? 1.0 : result < 0 ? 0.0 : 0.5;
if (u->val_scale && result != 0) {
int vp = u->val_points;
if (!vp) {
vp = board_size(b) - 1; vp *= vp; vp *= 2;
}
floating_t sval = (floating_t) abs(result) / vp;
sval = sval > 1 ? 1 : sval;
if (result < 0) sval = 1 - sval;
if (u->val_extra)
rval += u->val_scale * sval;
else
rval = (1 - u->val_scale) * rval + u->val_scale * sval;
// fprintf(stderr, "score %d => sval %f, rval %f\n", result, sval, rval);
}
return rval;
}
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 =)
}
