UCT: bestr_ratio 0.02 by default, best2_ratio 2.5 by default

[pachi.git] / uct / uct.c

#include <assert.h>
#include <math.h>
#include <pthread.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

#define DEBUG

#include "debug.h"
#include "board.h"
#include "gtp.h"
#include "move.h"
#include "mq.h"
#include "playout.h"
#include "playout/elo.h"
#include "playout/moggy.h"
#include "playout/light.h"
#include "random.h"
#include "timeinfo.h"
#include "tactics.h"
#include "uct/internal.h"
#include "uct/prior.h"
#include "uct/tree.h"
#include "uct/uct.h"
#include "uct/walk.h"

struct uct_policy *policy_ucb1_init(struct uct *u, char *arg);
struct uct_policy *policy_ucb1amaf_init(struct uct *u, char *arg);
static void uct_pondering_stop(struct uct *u);


/* Default number of simulations to perform per move.
 * Note that this is now in total over all threads! (Unless TM_ROOT.) */
#define MC_GAMES        80000
#define MC_GAMELEN      MAX_GAMELEN
static const struct time_info default_ti = {
        .period = TT_MOVE,
        .dim = TD_GAMES,
        .len = { .games = MC_GAMES },
};

/* How big proportion of ownermap counts must be of one color to consider
 * the point sure. */
#define GJ_THRES        0.8
/* How many games to consider at minimum before judging groups. */
#define GJ_MINGAMES     500

/* How often to inspect the tree from the main thread to check for playout
 * stop, progress reports, etc. (in seconds) */
#define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */

/* Once per how many simulations (per thread) to show a progress report line. */
#define TREE_SIMPROGRESS_INTERVAL 10000

/* When terminating uct_search() early, the safety margin to add to the
 * remaining playout number estimate when deciding whether the result can
 * still change. */
#define PLAYOUT_DELTA_SAFEMARGIN 1000


static void
setup_state(struct uct *u, struct board *b, enum stone color)
{
        u->t = tree_init(b, color, u->fast_alloc ? u->max_tree_size: 0);
        if (u->force_seed)
                fast_srandom(u->force_seed);
        if (UDEBUGL(0))
                fprintf(stderr, "Fresh board with random seed %lu\n", fast_getseed());
        //board_print(b, stderr);
        if (!u->no_book && b->moves == 0) {
                assert(color == S_BLACK);
                tree_load(u->t, b);
        }
}

static void
reset_state(struct uct *u)
{
        assert(u->t);
        tree_done(u->t); u->t = NULL;
}

static void
setup_dynkomi(struct uct *u, struct board *b, enum stone to_play)
{
        if (u->dynkomi > b->moves && u->t->use_extra_komi)
                u->t->extra_komi = uct_get_extra_komi(u, b);
        else
                u->t->extra_komi = 0;
}

static void
prepare_move(struct engine *e, struct board *b, enum stone color)
{
        struct uct *u = e->data;

        if (u->t) {
                /* Verify that we have sane state. */
                assert(b->es == u);
                assert(u->t && b->moves);
                if (color != stone_other(u->t->root_color)) {
                        fprintf(stderr, "Fatal: Non-alternating play detected %d %d\n",
                                color, u->t->root_color);
                        exit(1);
                }

        } else {
                /* We need fresh state. */
                b->es = u;
                setup_state(u, b, color);
        }

        u->ownermap.playouts = 0;
        memset(u->ownermap.map, 0, board_size2(b) * sizeof(u->ownermap.map[0]));
}

static void
dead_group_list(struct uct *u, struct board *b, struct move_queue *mq)
{
        struct group_judgement gj;
        gj.thres = GJ_THRES;
        gj.gs = alloca(board_size2(b) * sizeof(gj.gs[0]));
        board_ownermap_judge_group(b, &u->ownermap, &gj);
        groups_of_status(b, &gj, GS_DEAD, mq);
}

bool
uct_pass_is_safe(struct uct *u, struct board *b, enum stone color, bool pass_all_alive)
{
        if (u->ownermap.playouts < GJ_MINGAMES)
                return false;

        struct move_queue mq = { .moves = 0 };
        if (!pass_all_alive)
                dead_group_list(u, b, &mq);
        return pass_is_safe(b, color, &mq);
}


static void
uct_printhook_ownermap(struct board *board, coord_t c, FILE *f)
{
        struct uct *u = board->es;
        assert(u);
        const char chr[] = ":XO,"; // dame, black, white, unclear
        const char chm[] = ":xo,";
        char ch = chr[board_ownermap_judge_point(&u->ownermap, c, GJ_THRES)];
        if (ch == ',') { // less precise estimate then?
                ch = chm[board_ownermap_judge_point(&u->ownermap, c, 0.67)];
        }
        fprintf(f, "%c ", ch);
}

static char *
uct_notify_play(struct engine *e, struct board *b, struct move *m)
{
        struct uct *u = e->data;
        if (!u->t) {
                /* No state, create one - this is probably game beginning
                 * and we need to load the opening book right now. */
                prepare_move(e, b, m->color);
                assert(u->t);
        }

        /* Stop pondering. */
        /* XXX: If we are about to receive multiple 'play' commands,
         * e.g. in a rengo, we will not ponder during the rest of them. */
        uct_pondering_stop(u);

        if (is_resign(m->coord)) {
                /* Reset state. */
                reset_state(u);
                return NULL;
        }

        /* Promote node of the appropriate move to the tree root. */
        assert(u->t->root);
        if (!tree_promote_at(u->t, b, m->coord)) {
                if (UDEBUGL(0))
                        fprintf(stderr, "Warning: Cannot promote move node! Several play commands in row?\n");
                reset_state(u);
                return NULL;
        }
        /* Setting up dynkomi is not necessary here, probably, but we
         * better do it anyway for consistency reasons. */
        setup_dynkomi(u, b, stone_other(m->color));
        return NULL;
}

static char *
uct_chat(struct engine *e, struct board *b, char *cmd)
{
        struct uct *u = e->data;
        static char reply[1024];

        cmd += strspn(cmd, " \n\t");
        if (!strncasecmp(cmd, "winrate", 7)) {
                if (!u->t)
                        return "no game context (yet?)";
                enum stone color = u->t->root_color;
                struct tree_node *n = u->t->root;
                snprintf(reply, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
                         n->u.playouts, u->threads, stone2str(color), coord2sstr(n->coord, b),
                         tree_node_get_value(u->t, -1, n->u.value) * 100);
                if (u->t->use_extra_komi && abs(u->t->extra_komi) >= 0.5) {
                        sprintf(reply + strlen(reply), ", while self-imposing extra komi %.1f",
                                u->t->extra_komi);
                }
                strcat(reply, ".");
                return reply;
        }
        return NULL;
}

static void
uct_dead_group_list(struct engine *e, struct board *b, struct move_queue *mq)
{
        struct uct *u = e->data;

        /* This means the game is probably over, no use pondering on. */
        uct_pondering_stop(u);

        if (u->pass_all_alive)
                return; // no dead groups

        bool mock_state = false;

        if (!u->t) {
                /* No state, but we cannot just back out - we might
                 * have passed earlier, only assuming some stones are
                 * dead, and then re-connected, only to lose counting
                 * when all stones are assumed alive. */
                /* Mock up some state and seed the ownermap by few
                 * simulations. */
                prepare_move(e, b, S_BLACK); assert(u->t);
                for (int i = 0; i < GJ_MINGAMES; i++)
                        uct_playout(u, b, S_BLACK, u->t);
                mock_state = true;
        }

        dead_group_list(u, b, mq);

        if (mock_state) {
                /* Clean up the mock state in case we will receive
                 * a genmove; we could get a non-alternating-move
                 * error from prepare_move() in that case otherwise. */
                reset_state(u);
        }
}

static void
playout_policy_done(struct playout_policy *p)
{
        if (p->done) p->done(p);
        if (p->data) free(p->data);
        free(p);
}

static void
uct_done(struct engine *e)
{
        /* This is called on engine reset, especially when clear_board
         * is received and new game should begin. */
        struct uct *u = e->data;
        uct_pondering_stop(u);
        if (u->t) reset_state(u);
        free(u->ownermap.map);

        free(u->policy);
        free(u->random_policy);
        playout_policy_done(u->playout);
        uct_prior_done(u->prior);
}


/* Pachi threading structure (if uct_playouts_parallel() is used):
 *
 * main thread
 *   |         main(), GTP communication, ...
 *   |         starts and stops the search managed by thread_manager
 *   |
 * thread_manager
 *   |         spawns and collects worker threads
 *   |
 * worker0
 * worker1
 * ...
 * workerK
 *             uct_playouts() loop, doing descend-playout until uct_halt
 *
 * Another way to look at it is by functions (lines denote thread boundaries):
 *
 * | uct_genmove()
 * | uct_search()            (uct_search_start() .. uct_search_stop())
 * | -----------------------
 * | spawn_thread_manager()
 * | -----------------------
 * | spawn_worker()
 * V uct_playouts() */

/* Set in thread manager in case the workers should stop. */
volatile sig_atomic_t uct_halt = 0;
/* ID of the running worker thread. */
__thread int thread_id = -1;
/* ID of the thread manager. */
static pthread_t thread_manager;
static bool thread_manager_running;

static pthread_mutex_t finish_mutex = PTHREAD_MUTEX_INITIALIZER;
static pthread_cond_t finish_cond = PTHREAD_COND_INITIALIZER;
static volatile int finish_thread;
static pthread_mutex_t finish_serializer = PTHREAD_MUTEX_INITIALIZER;

struct spawn_ctx {
        int tid;
        struct uct *u;
        struct board *b;
        enum stone color;
        struct tree *t;
        unsigned long seed;
        int games;
};

static void *
spawn_worker(void *ctx_)
{
        struct spawn_ctx *ctx = ctx_;
        /* Setup */
        fast_srandom(ctx->seed);
        thread_id = ctx->tid;
        /* Run */
        ctx->games = uct_playouts(ctx->u, ctx->b, ctx->color, ctx->t);
        /* Finish */
        pthread_mutex_lock(&finish_serializer);
        pthread_mutex_lock(&finish_mutex);
        finish_thread = ctx->tid;
        pthread_cond_signal(&finish_cond);
        pthread_mutex_unlock(&finish_mutex);
        return ctx;
}

/* Thread manager, controlling worker threads. It must be called with
 * finish_mutex lock held, but it will unlock it itself before exiting;
 * this is necessary to be completely deadlock-free. */
/* The finish_cond can be signalled for it to stop; in that case,
 * the caller should set finish_thread = -1. */
/* After it is started, it will update mctx->t to point at some tree
 * used for the actual search (matters only for TM_ROOT), on return
 * it will set mctx->games to the number of performed simulations. */
static void *
spawn_thread_manager(void *ctx_)
{
        /* In thread_manager, we use only some of the ctx fields. */
        struct spawn_ctx *mctx = ctx_;
        struct uct *u = mctx->u;
        struct tree *t = mctx->t;
        bool shared_tree = u->parallel_tree;
        fast_srandom(mctx->seed);

        int played_games = 0;
        pthread_t threads[u->threads];
        int joined = 0;

        uct_halt = 0;

        /* Spawn threads... */
        for (int ti = 0; ti < u->threads; ti++) {
                struct spawn_ctx *ctx = malloc(sizeof(*ctx));
                ctx->u = u; ctx->b = mctx->b; ctx->color = mctx->color;
                mctx->t = ctx->t = shared_tree ? t : tree_copy(t);
                ctx->tid = ti; ctx->seed = fast_random(65536) + ti;
                pthread_create(&threads[ti], NULL, spawn_worker, ctx);
                if (UDEBUGL(2))
                        fprintf(stderr, "Spawned worker %d\n", ti);
        }

        /* ...and collect them back: */
        while (joined < u->threads) {
                /* Wait for some thread to finish... */
                pthread_cond_wait(&finish_cond, &finish_mutex);
                if (finish_thread < 0) {
                        /* Stop-by-caller. Tell the workers to wrap up. */
                        uct_halt = 1;
                        continue;
                }
                /* ...and gather its remnants. */
                struct spawn_ctx *ctx;
                pthread_join(threads[finish_thread], (void **) &ctx);
                played_games += ctx->games;
                joined++;
                if (!shared_tree) {
                        if (ctx->t == mctx->t) mctx->t = t;
                        tree_merge(t, ctx->t);
                        tree_done(ctx->t);
                }
                free(ctx);
                if (UDEBUGL(2))
                        fprintf(stderr, "Joined worker %d\n", finish_thread);
                pthread_mutex_unlock(&finish_serializer);
        }

        pthread_mutex_unlock(&finish_mutex);

        if (!shared_tree)
                tree_normalize(mctx->t, u->threads);

        mctx->games = played_games;
        return mctx;
}

static struct spawn_ctx *
uct_search_start(struct uct *u, struct board *b, enum stone color, struct tree *t)
{
        assert(u->threads > 0);
        assert(!thread_manager_running);

        struct spawn_ctx ctx = { .u = u, .b = b, .color = color, .t = t, .seed = fast_random(65536) };
        static struct spawn_ctx mctx; mctx = ctx;
        pthread_mutex_lock(&finish_mutex);
        pthread_create(&thread_manager, NULL, spawn_thread_manager, &mctx);
        thread_manager_running = true;
        return &mctx;
}

static struct spawn_ctx *
uct_search_stop(void)
{
        assert(thread_manager_running);

        /* Signal thread manager to stop the workers. */
        pthread_mutex_lock(&finish_mutex);
        finish_thread = -1;
        pthread_cond_signal(&finish_cond);
        pthread_mutex_unlock(&finish_mutex);

        /* Collect the thread manager. */
        struct spawn_ctx *pctx;
        thread_manager_running = false;
        pthread_join(thread_manager, (void **) &pctx);
        return pctx;
}


/* Determine whether we should terminate the search early. */
static bool
uct_search_stop_early(struct uct *u, struct tree *t, struct board *b,
                struct time_info *ti, struct time_stop *stop,
                struct tree_node *best, struct tree_node *best2,
                int base_playouts, int i)
{
        /* Early break in won situation. */
        if (best->u.playouts >= 2000 && tree_node_get_value(t, 1, best->u.value) >= u->loss_threshold)
                return true;
        /* Earlier break in super-won situation. */
        if (best->u.playouts >= 500 && tree_node_get_value(t, 1, best->u.value) >= 0.95)
                return true;

        /* Break early if we estimate the second-best move cannot
         * catch up in assigned time anymore. We use all our time
         * if we are in byoyomi with single stone remaining in our
         * period, however - it's better to pre-ponder. */
        bool time_indulgent = (!ti->len.t.main_time && ti->len.t.byoyomi_stones == 1);
        if (best2 && ti->dim == TD_WALLTIME && !time_indulgent) {
                double elapsed = time_now() - ti->len.t.timer_start;
                double remaining = stop->worst.time - elapsed;
                double pps = ((double)i - base_playouts) / elapsed;
                double estplayouts = remaining * pps + PLAYOUT_DELTA_SAFEMARGIN;
                if (best->u.playouts > best2->u.playouts + estplayouts) {
                        if (UDEBUGL(2))
                                fprintf(stderr, "Early stop, result cannot change: "
                                        "best %d, best2 %d, estimated %f simulations to go\n",
                                        best->u.playouts, best2->u.playouts, estplayouts);
                        return true;
                }
        }

        return false;
}

/* Determine whether we should terminate the search later. */
static bool
uct_search_keep_looking(struct uct *u, struct tree *t, struct board *b,
                struct tree_node *best, struct tree_node *best2,
                struct tree_node *bestr, struct tree_node *winner, int i)
{
        if (!best) {
                if (UDEBUGL(2))
                        fprintf(stderr, "Did not find best move, still trying...\n");
                return true;
        }

        if (u->best2_ratio > 0) {
                /* Check best/best2 simulations ratio. If the
                 * two best moves give very similar results,
                 * keep simulating. */
                if (best2 && best2->u.playouts
                    && (double)best->u.playouts / best2->u.playouts < u->best2_ratio) {
                        if (UDEBUGL(2))
                                fprintf(stderr, "Best2 ratio %f < threshold %f\n",
                                        (double)best->u.playouts / best2->u.playouts,
                                        u->best2_ratio);
                        return true;
                }
        }

        if (u->bestr_ratio > 0) {
                /* Check best, best_best value difference. If the best move
                 * and its best child do not give similar enough results,
                 * keep simulating. */
                if (bestr && bestr->u.playouts
                    && fabs((double)best->u.value - bestr->u.value) > u->bestr_ratio) {
                        if (UDEBUGL(2))
                                fprintf(stderr, "Bestr delta %f > threshold %f\n",
                                        fabs((double)best->u.value - bestr->u.value),
                                        u->bestr_ratio);
                        return true;
                }
        }

        if (winner && winner != best) {
                /* Keep simulating if best explored
                 * does not have also highest value. */
                if (UDEBUGL(2))
                        fprintf(stderr, "[%d] best %3s [%d] %f != winner %3s [%d] %f\n", i,
                                coord2sstr(best->coord, t->board),
                                best->u.playouts, tree_node_get_value(t, 1, best->u.value),
                                coord2sstr(winner->coord, t->board),
                                winner->u.playouts, tree_node_get_value(t, 1, winner->u.value));
                return true;
        }

        /* No reason to keep simulating, bye. */
        return false;
}

/* Run time-limited MCTS search on foreground. */
static int
uct_search(struct uct *u, struct board *b, struct time_info *ti, enum stone color, struct tree *t)
{
        int base_playouts = u->t->root->u.playouts;
        if (UDEBUGL(2) && base_playouts > 0)
                fprintf(stderr, "<pre-simulated %d games skipped>\n", base_playouts);

        /* Set up time conditions. */
        if (ti->period == TT_NULL) *ti = default_ti;
        struct time_stop stop;
        time_stop_conditions(ti, b, u->fuseki_end, u->yose_start, &stop);

        /* Number of last game with progress print. */
        int last_print = t->root->u.playouts;
        /* Number of simulations to wait before next print. */
        int print_interval = TREE_SIMPROGRESS_INTERVAL * (u->thread_model == TM_ROOT ? 1 : u->threads);
        /* Printed notification about full memory? */
        bool print_fullmem = false;

        struct spawn_ctx *ctx = uct_search_start(u, b, color, t);

        /* The search tree is ctx->t. This is normally == t, but in case of
         * TM_ROOT, it is one of the trees belonging to the independent
         * workers. It is important to reference ctx->t directly since the
         * thread manager will swap the tree pointer asynchronously. */
        /* XXX: This means TM_ROOT support is suboptimal since single stalled
         * thread can stall the others in case of limiting the search by game
         * count. However, TM_ROOT just does not deserve any more extra code
         * right now. */

        struct tree_node *best = NULL;
        struct tree_node *best2 = NULL; // Second-best move.
        struct tree_node *bestr = NULL; // best's best child.
        struct tree_node *winner = NULL;

        double busywait_interval = TREE_BUSYWAIT_INTERVAL;

        /* Now, just periodically poll the search tree. */
        while (1) {
                time_sleep(busywait_interval);
                /* busywait_interval should never be less than desired time, or the
                 * time control is broken. But if it happens to be less, we still search
                 * at least 100ms otherwise the move is completely random. */

                int i = ctx->t->root->u.playouts;

                /* Print progress? */
                if (i - last_print > print_interval) {
                        last_print += print_interval; // keep the numbers tidy
                        uct_progress_status(u, ctx->t, color, last_print);
                }
                if (!print_fullmem && ctx->t->nodes_size > u->max_tree_size) {
                        if (UDEBUGL(2))
                                fprintf(stderr, "memory limit hit (%ld > %lu)\n", ctx->t->nodes_size, u->max_tree_size);
                        print_fullmem = true;
                }

                best = u->policy->choose(u->policy, ctx->t->root, b, color, resign);
                if (best) best2 = u->policy->choose(u->policy, ctx->t->root, b, color, best->coord);

                /* Possibly stop search early if it's no use to try on. */
                if (best && uct_search_stop_early(u, ctx->t, b, ti, &stop, best, best2, base_playouts, i))
                        break;

                /* Check against time settings. */
                bool desired_done = false;
                if (ti->dim == TD_WALLTIME) {
                        double elapsed = time_now() - ti->len.t.timer_start;
                        if (elapsed > stop.worst.time) break;
                        desired_done = elapsed > stop.desired.time;

                } else { assert(ti->dim == TD_GAMES);
                        if (i > stop.worst.playouts) break;
                        desired_done = i > stop.desired.playouts;
                }

                /* We want to stop simulating, but are willing to keep trying
                 * if we aren't completely sure about the winner yet. */
                if (desired_done) {
                        if (u->policy->winner && u->policy->evaluate)
                                winner = u->policy->winner(u->policy, ctx->t, ctx->t->root);
                        if (best)
                                bestr = u->policy->choose(u->policy, best, b, stone_other(color), resign);
                        if (!uct_search_keep_looking(u, ctx->t, b, best, best2, bestr, winner, i))
                                break;
                }

                /* TODO: Early break if best->variance goes under threshold and we already
                 * have enough playouts (possibly thanks to book or to pondering)? */
        }

        ctx = uct_search_stop();

        if (UDEBUGL(2))
                tree_dump(t, u->dumpthres);
        if (UDEBUGL(0))
                uct_progress_status(u, t, color, ctx->games);

        return ctx->games;
}


/* Start pondering background with @color to play. */
static void
uct_pondering_start(struct uct *u, struct board *b0, struct tree *t, enum stone color)
{
        if (UDEBUGL(1))
                fprintf(stderr, "Starting to ponder with color %s\n", stone2str(stone_other(color)));
        u->pondering = true;

        /* We need a local board copy to ponder upon. */
        struct board *b = malloc(sizeof(*b)); board_copy(b, b0);

        /* *b0 did not have the genmove'd move played yet. */
        struct move m = { t->root->coord, t->root_color };
        int res = board_play(b, &m);
        assert(res >= 0);
        setup_dynkomi(u, b, stone_other(m.color));

        /* Start MCTS manager thread "headless". */
        uct_search_start(u, b, color, t);
}

/* uct_search_stop() frontend for the pondering (non-genmove) mode. */
static void
uct_pondering_stop(struct uct *u)
{
        u->pondering = false;
        if (!thread_manager_running)
                return;

        /* Stop the thread manager. */
        struct spawn_ctx *ctx = uct_search_stop();
        if (UDEBUGL(1)) {
                fprintf(stderr, "(pondering) ");
                uct_progress_status(u, ctx->t, ctx->color, ctx->games);
        }
        free(ctx->b);
}


static coord_t *
uct_genmove(struct engine *e, struct board *b, struct time_info *ti, enum stone color, bool pass_all_alive)
{
        double start_time = time_now();
        struct uct *u = e->data;

        if (b->superko_violation) {
                fprintf(stderr, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
                fprintf(stderr, "Maybe you play with situational instead of positional superko?\n");
                fprintf(stderr, "I'm going to ignore the violation, but note that I may miss\n");
                fprintf(stderr, "some moves valid under this ruleset because of this.\n");
                b->superko_violation = false;
        }

        /* Seed the tree. */
        uct_pondering_stop(u);
        prepare_move(e, b, color);
        assert(u->t);

        /* How to decide whether to use dynkomi in this game? Since we use
         * pondering, it's not simple "who-to-play" matter. Decide based on
         * the last genmove issued. */
        u->t->use_extra_komi = !!(u->dynkomi_mask & color);
        setup_dynkomi(u, b, color);

        /* Make pessimistic assumption about komi for Japanese rules to
         * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
         * The rules usually give the same winner if the integer part of komi
         * is odd so we adjust the komi only if it is even (for a board of
         * odd size). We are not trying  to get an exact evaluation for rare
         * cases of seki. For details see http://home.snafu.de/jasiek/parity.html
         * TODO: Support the kgs-rules command once available. */
        if (u->territory_scoring && (((int)floor(b->komi) + b->size) & 1)) {
                b->komi += (color == S_BLACK ? 1.0 : -1.0);
                if (UDEBUGL(0))
                        fprintf(stderr, "Setting komi to %.1f assuming Japanese rules\n",
                                b->komi);
        }

        /* Perform the Monte Carlo Tree Search! */
        int played_games = uct_search(u, b, ti, color, u->t);

        /* Choose the best move from the tree. */
        struct tree_node *best = u->policy->choose(u->policy, u->t->root, b, color, resign);
        if (!best) {
                reset_state(u);
                return coord_copy(pass);
        }
        if (UDEBUGL(1))
                fprintf(stderr, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n",
                        coord2sstr(best->coord, b), coord_x(best->coord, b), coord_y(best->coord, b),
                        tree_node_get_value(u->t, 1, best->u.value),
                        best->u.playouts, u->t->root->u.playouts, played_games);

        /* Do not resign if we're so short of time that evaluation of best move is completely
         * unreliable, we might be winning actually. In this case best is almost random but
         * still better than resign. */
        if (tree_node_get_value(u->t, 1, best->u.value) < u->resign_ratio && !is_pass(best->coord)
            && best->u.playouts > GJ_MINGAMES) {
                reset_state(u);
                return coord_copy(resign);
        }

        /* If the opponent just passed and we win counting, always
         * pass as well. */
        if (b->moves > 1 && is_pass(b->last_move.coord)) {
                /* Make sure enough playouts are simulated. */
                while (u->ownermap.playouts < GJ_MINGAMES)
                        uct_playout(u, b, color, u->t);
                if (uct_pass_is_safe(u, b, color, u->pass_all_alive || pass_all_alive)) {
                        if (UDEBUGL(0))
                                fprintf(stderr, "<Will rather pass, looks safe enough.>\n");
                        best->coord = pass;
                }
        }

        tree_promote_node(u->t, &best);
        /* After a pass, pondering is harmful for two reasons:
         * (i) We might keep pondering even when the game is over.
         * Of course this is the case for opponent resign as well.
         * (ii) More importantly, the ownermap will get skewed since
         * the UCT will start cutting off any playouts. */
        if (u->pondering_opt && !is_pass(best->coord)) {
                uct_pondering_start(u, b, u->t, stone_other(color));
        }
        if (UDEBUGL(2)) {
                double time = time_now() - start_time + 0.000001; /* avoid divide by zero */
                fprintf(stderr, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
                        time, (int)(played_games/time), (int)(played_games/time/u->threads));
        }
        return coord_copy(best->coord);
}


bool
uct_genbook(struct engine *e, struct board *b, struct time_info *ti, enum stone color)
{
        struct uct *u = e->data;
        if (!u->t) prepare_move(e, b, color);
        assert(u->t);

        if (ti->dim == TD_GAMES) {
                /* Don't count in games that already went into the book. */
                ti->len.games += u->t->root->u.playouts;
        }
        uct_search(u, b, ti, color, u->t);

        assert(ti->dim == TD_GAMES);
        tree_save(u->t, b, ti->len.games / 100);

        return true;
}

void
uct_dumpbook(struct engine *e, struct board *b, enum stone color)
{
        struct uct *u = e->data;
        struct tree *t = tree_init(b, color, u->fast_alloc ? u->max_tree_size: 0);
        tree_load(t, b);
        tree_dump(t, 0);
        tree_done(t);
}


struct uct *
uct_state_init(char *arg, struct board *b)
{
        struct uct *u = calloc(1, sizeof(struct uct));

        u->debug_level = 3;
        u->gamelen = MC_GAMELEN;
        u->mercymin = 0;
        u->expand_p = 2;
        u->dumpthres = 1000;
        u->playout_amaf = true;
        u->playout_amaf_nakade = false;
        u->amaf_prior = false;
        u->max_tree_size = 3072ULL * 1048576;

        if (board_size(b) - 2 >= 19)
                u->dynkomi = 200;
        u->dynkomi_mask = S_BLACK;
        u->handicap_value = 7;

        u->threads = 1;
        u->thread_model = TM_TREEVL;
        u->parallel_tree = true;
        u->virtual_loss = true;

        u->fuseki_end = 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
        u->yose_start = 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
        u->bestr_ratio = 0.02;
        // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
        // TODO: Further tuning and experiments with better time allocation schemes.
        u->best2_ratio = 2.5;

        u->val_scale = 0.04; u->val_points = 40;

        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)
                                        u->debug_level = atoi(optval);
                                else
                                        u->debug_level++;
                        } else if (!strcasecmp(optname, "mercy") && optval) {
                                /* Minimal difference of black/white captures
                                 * to stop playout - "Mercy Rule". Speeds up
                                 * hopeless playouts at the expense of some
                                 * accuracy. */
                                u->mercymin = atoi(optval);
                        } else if (!strcasecmp(optname, "gamelen") && optval) {
                                u->gamelen = atoi(optval);
                        } else if (!strcasecmp(optname, "expand_p") && optval) {
                                u->expand_p = atoi(optval);
                        } else if (!strcasecmp(optname, "dumpthres") && optval) {
                                u->dumpthres = atoi(optval);
                        } else if (!strcasecmp(optname, "best2_ratio") && optval) {
                                /* If set, prolong simulating while
                                 * first_best/second_best playouts ratio
                                 * is less than best2_ratio. */
                                u->best2_ratio = atof(optval);
                        } else if (!strcasecmp(optname, "bestr_ratio") && optval) {
                                /* If set, prolong simulating while
                                 * best,best_best_child values delta
                                 * is more than bestr_ratio. */
                                u->bestr_ratio = atof(optval);
                        } else if (!strcasecmp(optname, "playout_amaf")) {
                                /* Whether to include random playout moves in
                                 * AMAF as well. (Otherwise, only tree moves
                                 * are included in AMAF. Of course makes sense
                                 * only in connection with an AMAF policy.) */
                                /* with-without: 55.5% (+-4.1) */
                                if (optval && *optval == '0')
                                        u->playout_amaf = false;
                                else
                                        u->playout_amaf = true;
                        } else if (!strcasecmp(optname, "playout_amaf_nakade")) {
                                /* Whether to include nakade moves from playouts
                                 * in the AMAF statistics; this tends to nullify
                                 * the playout_amaf effect by adding too much
                                 * noise. */
                                if (optval && *optval == '0')
                                        u->playout_amaf_nakade = false;
                                else
                                        u->playout_amaf_nakade = true;
                        } else if (!strcasecmp(optname, "playout_amaf_cutoff") && optval) {
                                /* Keep only first N% of playout stage AMAF
                                 * information. */
                                u->playout_amaf_cutoff = atoi(optval);
                        } else if ((!strcasecmp(optname, "policy") || !strcasecmp(optname, "random_policy")) && optval) {
                                char *policyarg = strchr(optval, ':');
                                struct uct_policy **p = !strcasecmp(optname, "policy") ? &u->policy : &u->random_policy;
                                if (policyarg)
                                        *policyarg++ = 0;
                                if (!strcasecmp(optval, "ucb1")) {
                                        *p = policy_ucb1_init(u, policyarg);
                                } else if (!strcasecmp(optval, "ucb1amaf")) {
                                        *p = policy_ucb1amaf_init(u, policyarg);
                                } else {
                                        fprintf(stderr, "UCT: Invalid tree policy %s\n", optval);
                                        exit(1);
                                }
                        } else if (!strcasecmp(optname, "playout") && optval) {
                                char *playoutarg = strchr(optval, ':');
                                if (playoutarg)
                                        *playoutarg++ = 0;
                                if (!strcasecmp(optval, "moggy")) {
                                        u->playout = playout_moggy_init(playoutarg, b);
                                } else if (!strcasecmp(optval, "light")) {
                                        u->playout = playout_light_init(playoutarg, b);
                                } else if (!strcasecmp(optval, "elo")) {
                                        u->playout = playout_elo_init(playoutarg, b);
                                } else {
                                        fprintf(stderr, "UCT: Invalid playout policy %s\n", optval);
                                        exit(1);
                                }
                        } else if (!strcasecmp(optname, "prior") && optval) {
                                u->prior = uct_prior_init(optval, b);
                        } else if (!strcasecmp(optname, "amaf_prior") && optval) {
                                u->amaf_prior = atoi(optval);
                        } else if (!strcasecmp(optname, "threads") && optval) {
                                /* By default, Pachi will run with only single
                                 * tree search thread! */
                                u->threads = atoi(optval);
                        } else if (!strcasecmp(optname, "thread_model") && optval) {
                                if (!strcasecmp(optval, "root")) {
                                        /* Root parallelization - each thread
                                         * does independent search, trees are
                                         * merged at the end. */
                                        u->thread_model = TM_ROOT;
                                        u->parallel_tree = false;
                                        u->virtual_loss = false;
                                } else if (!strcasecmp(optval, "tree")) {
                                        /* Tree parallelization - all threads
                                         * grind on the same tree. */
                                        u->thread_model = TM_TREE;
                                        u->parallel_tree = true;
                                        u->virtual_loss = false;
                                } else if (!strcasecmp(optval, "treevl")) {
                                        /* Tree parallelization, but also
                                         * with virtual losses - this discou-
                                         * rages most threads choosing the
                                         * same tree branches to read. */
                                        u->thread_model = TM_TREEVL;
                                        u->parallel_tree = true;
                                        u->virtual_loss = true;
                                } else {
                                        fprintf(stderr, "UCT: Invalid thread model %s\n", optval);
                                        exit(1);
                                }
                        } else if (!strcasecmp(optname, "pondering")) {
                                /* Keep searching even during opponent's turn. */
                                u->pondering_opt = !optval || atoi(optval);
                        } else if (!strcasecmp(optname, "fuseki_end") && optval) {
                                /* At the very beginning it's not worth thinking
                                 * too long because the playout evaluations are
                                 * very noisy. So gradually increase the thinking
                                 * time up to maximum when fuseki_end percent
                                 * of the board has been played.
                                 * This only applies if we are not in byoyomi. */
                                u->fuseki_end = atoi(optval);
                        } else if (!strcasecmp(optname, "yose_start") && optval) {
                                /* When yose_start percent of the board has been
                                 * played, or if we are in byoyomi, stop spending
                                 * more time and spread the remaining time
                                 * uniformly.
                                 * Between fuseki_end and yose_start, we spend
                                 * a constant proportion of the remaining time
                                 * on each move. (yose_start should actually
                                 * be much earlier than when real yose start,
                                 * but "yose" is a good short name to convey
                                 * the idea.) */
                                u->yose_start = atoi(optval);
                        } else if (!strcasecmp(optname, "force_seed") && optval) {
                                u->force_seed = atoi(optval);
                        } else if (!strcasecmp(optname, "no_book")) {
                                u->no_book = true;
                        } else if (!strcasecmp(optname, "dynkomi")) {
                                /* Dynamic komi in handicap game; linearly
                                 * decreases to basic settings until move
                                 * #optval. */
                                u->dynkomi = optval ? atoi(optval) : 200;
                        } else if (!strcasecmp(optname, "dynkomi_mask") && optval) {
                                /* Bitmask of colors the player must be
                                 * for dynkomi be applied; you may want
                                 * to use dynkomi_mask=3 to allow dynkomi
                                 * even in games where Pachi is white. */
                                u->dynkomi_mask = atoi(optval);
                        } else if (!strcasecmp(optname, "handicap_value") && optval) {
                                /* Point value of single handicap stone,
                                 * for dynkomi computation. */
                                u->handicap_value = atoi(optval);
                        } else if (!strcasecmp(optname, "val_scale") && optval) {
                                /* How much of the game result value should be
                                 * influenced by win size. Zero means it isn't. */
                                u->val_scale = atof(optval);
                        } else if (!strcasecmp(optname, "val_points") && optval) {
                                /* Maximum size of win to be scaled into game
                                 * result value. Zero means boardsize^2. */
                                u->val_points = atoi(optval) * 2; // result values are doubled
                        } else if (!strcasecmp(optname, "val_extra")) {
                                /* If false, the score coefficient will be simply
                                 * added to the value, instead of scaling the result
                                 * coefficient because of it. */
                                u->val_extra = !optval || atoi(optval);
                        } else if (!strcasecmp(optname, "root_heuristic") && optval) {
                                /* Whether to bias exploration by root node values
                                 * (must be supported by the used policy).
                                 * 0: Don't.
                                 * 1: Do, value = result.
                                 * Try to temper the result:
                                 * 2: Do, value = 0.5+(result-expected)/2.
                                 * 3: Do, value = 0.5+bzz((result-expected)^2). */
                                u->root_heuristic = atoi(optval);
                        } else if (!strcasecmp(optname, "pass_all_alive")) {
                                /* Whether to consider all stones alive at the game
                                 * end instead of marking dead groupd. */
                                u->pass_all_alive = !optval || atoi(optval);
                        } else if (!strcasecmp(optname, "territory_scoring")) {
                                /* Use territory scoring (default is area scoring).
                                 * An explicit kgs-rules command overrides this. */
                                u->territory_scoring = !optval || atoi(optval);
                        } else if (!strcasecmp(optname, "random_policy_chance") && optval) {
                                /* If specified (N), with probability 1/N, random_policy policy
                                 * descend is used instead of main policy descend; useful
                                 * if specified policy (e.g. UCB1AMAF) can make unduly biased
                                 * choices sometimes, you can fall back to e.g.
                                 * random_policy=UCB1. */
                                u->random_policy_chance = atoi(optval);
                        } else if (!strcasecmp(optname, "max_tree_size") && optval) {
                                /* Maximum amount of memory [MiB] consumed by the move tree.
                                 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
                                 * this limits size of only one of the trees, not all of them
                                 * together. */
                                u->max_tree_size = atol(optval) * 1048576;
                        } else if (!strcasecmp(optname, "fast_alloc")) {
                                u->fast_alloc = !optval || atoi(optval);
                        } else if (!strcasecmp(optname, "banner") && optval) {
                                /* Additional banner string. This must come as the
                                 * last engine parameter. */
                                if (*next) *--next = ',';
                                u->banner = strdup(optval);
                                break;
                        } else {
                                fprintf(stderr, "uct: Invalid engine argument %s or missing value\n", optname);
                                exit(1);
                        }
                }
        }

        u->resign_ratio = 0.2; /* Resign when most games are lost. */
        u->loss_threshold = 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
        if (!u->policy)
                u->policy = policy_ucb1amaf_init(u, NULL);

        if (!!u->random_policy_chance ^ !!u->random_policy) {
                fprintf(stderr, "uct: Only one of random_policy and random_policy_chance is set\n");
                exit(1);
        }

        if (u->fast_alloc && !u->parallel_tree) {
                fprintf(stderr, "fast_alloc not supported with root parallelization.\n");
                exit(1);
        }

        if (!u->prior)
                u->prior = uct_prior_init(NULL, b);

        if (!u->playout)
                u->playout = playout_moggy_init(NULL, b);
        u->playout->debug_level = u->debug_level;

        u->ownermap.map = malloc(board_size2(b) * sizeof(u->ownermap.map[0]));

        /* Some things remain uninitialized for now - the opening book
         * is not loaded and the tree not set up. */
        /* This will be initialized in setup_state() at the first move
         * received/requested. This is because right now we are not aware
         * about any komi or handicap setup and such. */

        return u;
}

struct engine *
engine_uct_init(char *arg, struct board *b)
{
        struct uct *u = uct_state_init(arg, b);
        struct engine *e = calloc(1, sizeof(struct engine));
        e->name = "UCT Engine";
        e->printhook = uct_printhook_ownermap;
        e->notify_play = uct_notify_play;
        e->chat = uct_chat;
        e->genmove = uct_genmove;
        e->dead_group_list = uct_dead_group_list;
        e->done = uct_done;
        e->data = u;

        const char banner[] = "I'm playing UCT. When I'm losing, I will resign, "
                "if I think I win, I play until you pass. "
                "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
        if (!u->banner) u->banner = "";
        e->comment = malloc(sizeof(banner) + strlen(u->banner) + 1);
        sprintf(e->comment, "%s %s", banner, u->banner);

        return e;
}