pachi_py/pachi/uct/prior.c (325 lines of code) (raw):
#include <assert.h>
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define DEBUG
#include "board.h"
#include "debug.h"
#include "joseki/base.h"
#include "move.h"
#include "random.h"
#include "dcnn.h"
#include "tactics/ladder.h"
#include "tactics/util.h"
#include "uct/internal.h"
#include "uct/plugins.h"
#include "uct/prior.h"
#include "uct/tree.h"
/* Applying heuristic values to the tree nodes, skewing the reading in
* most interesting directions. */
/* TODO: Introduce foreach_fpoint() to iterate only over non-occupied
* positions. */
struct uct_prior {
/* Equivalent experience for prior knowledge. MoGo paper recommends
* 50 playouts per source; in practice, esp. with RAVE, about 6
* playouts per source seems best. */
int eqex;
int even_eqex, policy_eqex, b19_eqex, eye_eqex, ko_eqex, plugin_eqex, joseki_eqex, pattern_eqex;
int dcnn_eqex;
int cfgdn; int *cfgd_eqex;
bool prune_ladders;
};
void
uct_prior_even(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{even} */
/* This may be dubious for normal UCB1 but is essential for
* reading stability of RAVE, it appears. */
add_prior_value(map, pass, 0.5, u->prior->even_eqex);
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
add_prior_value(map, c, 0.5, u->prior->even_eqex);
} foreach_free_point_end;
}
void
uct_prior_eye(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Discourage playing into our own eyes. However, we cannot
* completely prohibit it:
* #######
* ...XX.#
* XOOOXX#
* X.OOOO#
* .XXXX.# */
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
if (!board_is_one_point_eye(map->b, c, map->to_play))
continue;
add_prior_value(map, c, 0, u->prior->eye_eqex);
} foreach_free_point_end;
}
#ifdef DCNN
#define DCNN_BEST_N 5
static void
find_dcnn_best_moves(struct prior_map *map, float *r, coord_t *best, float *best_r)
{
struct board *b = map->b;
for (int i = 0; i < DCNN_BEST_N; i++)
best[i] = pass;
foreach_free_point(b) {
if (!map->consider[c])
continue;
int k = (coord_x(c, b) - 1) * 19 + (coord_y(c, b) - 1);
for (int i = 0; i < DCNN_BEST_N; i++)
if (r[k] > best_r[i]) {
for (int j = DCNN_BEST_N - 1; j > i; j--) { // shift
best_r[j] = best_r[j - 1];
best[j] = best[j - 1];
}
best_r[i] = r[k];
best[i] = c;
break;
}
} foreach_free_point_end;
}
static void
print_dcnn_best_moves(struct tree_node *node, struct prior_map *map,
coord_t *best, float *best_r)
{
fprintf(stderr, "dcnn best: [ ");
for (int i = 0; i < DCNN_BEST_N; i++)
fprintf(stderr, "%s ", coord2sstr(best[i], map->b));
fprintf(stderr, "] ");
fprintf(stderr, "[ ");
for (int i = 0; i < DCNN_BEST_N; i++)
fprintf(stderr, "%.2f ", best_r[i]);
fprintf(stderr, "]\n");
}
static void
uct_prior_dcnn(struct uct *u, struct tree_node *node, struct prior_map *map)
{
float r[19 * 19];
float best_r[DCNN_BEST_N] = { 0.0, };
coord_t best_moves[DCNN_BEST_N];
dcnn_get_moves(map->b, map->to_play, r);
find_dcnn_best_moves(map, r, best_moves, best_r);
print_dcnn_best_moves(node, map, best_moves, best_r);
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
int i = coord_x(c, map->b) - 1;
int j = coord_y(c, map->b) - 1;
assert(i >= 0 && i < 19);
assert(j >= 0 && j < 19);
float val = r[i * 19 + j];
if (isnan(val) || val < 0.001)
continue;
assert(val >= 0.0 && val <= 1.0);
add_prior_value(map, c, 1, sqrt(val) * u->prior->dcnn_eqex);
} foreach_free_point_end;
}
#else
#define uct_prior_dcnn(u, node, map)
#endif /* DCNN */
void
uct_prior_ko(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Favor fighting ko, if we took it le 10 moves ago. */
coord_t ko = map->b->last_ko.coord;
if (is_pass(ko) || map->b->moves - map->b->last_ko_age > 10 || !map->consider[ko])
return;
// fprintf(stderr, "prior ko-fight @ %s %s\n", stone2str(map->to_play), coord2sstr(ko, map->b));
add_prior_value(map, ko, 1, u->prior->ko_eqex);
}
void
uct_prior_b19(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{b19} */
/* Specific hints for 19x19 board - priors for certain edge distances. */
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
int d = coord_edge_distance(c, map->b);
if (d != 0 && d != 2)
continue;
/* The bonus applies only with no stones in immediate
* vincinity. */
if (board_stone_radar(map->b, c, 2))
continue;
/* First line: 0 */
/* Third line: 1 */
add_prior_value(map, c, d == 2, u->prior->b19_eqex);
} foreach_free_point_end;
}
void
uct_prior_playout(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{playout-policy} */
if (u->playout->assess)
u->playout->assess(u->playout, map, u->prior->policy_eqex);
}
void
uct_prior_cfgd(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{common_fate_graph_distance} */
/* Give bonus to moves local to the last move, where "local" means
* local in terms of groups, not just manhattan distance. */
if (is_pass(map->b->last_move.coord) || is_resign(map->b->last_move.coord))
return;
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
if (map->distances[c] > u->prior->cfgdn)
continue;
assert(map->distances[c] != 0);
int bonus = u->prior->cfgd_eqex[map->distances[c]];
add_prior_value(map, c, 1, bonus);
} foreach_free_point_end;
}
void
uct_prior_joseki(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{joseki} */
if (!u->jdict)
return;
for (int i = 0; i < 4; i++) {
hash_t h = map->b->qhash[i] & joseki_hash_mask;
coord_t *cc = u->jdict->patterns[h].moves[map->to_play - 1];
if (!cc) continue;
for (; !is_pass(*cc); cc++) {
if (coord_quadrant(*cc, map->b) != i)
continue;
add_prior_value(map, *cc, 1.0, u->prior->joseki_eqex);
}
}
}
void
uct_prior_pattern(struct uct *u, struct tree_node *node, struct prior_map *map)
{
/* Q_{pattern} */
if (!u->pat.pd)
return;
struct board *b = map->b;
struct pattern pats[b->flen];
floating_t probs[b->flen];
pattern_rate_moves(&u->pat, b, map->to_play, pats, probs);
if (UDEBUGL(5)) {
fprintf(stderr, "Pattern prior at node %s\n", coord2sstr(node->coord, b));
board_print(b, stderr);
}
for (int f = 0; f < b->flen; f++) {
if (isnan(probs[f]) || probs[f] < 0.001)
continue;
assert(!is_pass(b->f[f]));
if (UDEBUGL(5)) {
char s[256]; pattern2str(s, &pats[f]);
fprintf(stderr, "\t%s: %.3f %s\n", coord2sstr(b->f[f], b), probs[f], s);
}
add_prior_value(map, b->f[f], 1.0, sqrt(probs[f]) * u->prior->pattern_eqex);
}
}
void
uct_prior(struct uct *u, struct tree_node *node, struct prior_map *map)
{
if (u->prior->prune_ladders && !board_playing_ko_threat(map->b)) {
foreach_free_point(map->b) {
if (!map->consider[c])
continue;
group_t atari_neighbor = board_get_atari_neighbor(map->b, c, map->to_play);
if (atari_neighbor && is_ladder(map->b, c, atari_neighbor, true)) {
if (UDEBUGL(5))
fprintf(stderr, "Pruning ladder move %s\n", coord2sstr(c, map->b));
map->consider[c] = false;
}
} foreach_free_point_end;
}
if (u->prior->even_eqex)
uct_prior_even(u, node, map);
if (u->prior->eye_eqex)
uct_prior_eye(u, node, map);
if (u->prior->ko_eqex)
uct_prior_ko(u, node, map);
if (u->prior->b19_eqex)
uct_prior_b19(u, node, map);
if (!node->parent) // Use dcnn for root priors
if (u->prior->dcnn_eqex)
uct_prior_dcnn(u, node, map);
if (u->prior->policy_eqex)
uct_prior_playout(u, node, map);
if (u->prior->cfgd_eqex)
uct_prior_cfgd(u, node, map);
if (u->prior->joseki_eqex)
uct_prior_joseki(u, node, map);
if (u->prior->pattern_eqex)
uct_prior_pattern(u, node, map);
if (u->prior->plugin_eqex)
plugin_prior(u->plugins, node, map, u->prior->plugin_eqex);
}
struct uct_prior *
uct_prior_init(char *arg, struct board *b, struct uct *u)
{
struct uct_prior *p = calloc2(1, sizeof(struct uct_prior));
p->even_eqex = p->policy_eqex = p->b19_eqex = p->eye_eqex = p->ko_eqex = p->plugin_eqex = -100;
/* FIXME: Optimal pattern_eqex is about -1000 with small playout counts
* but only -400 on a cluster. We need a better way to set the default
* here. */
p->pattern_eqex = -800;
/* Best value for dcnn_eqex so far seems to be 1300 with ~88% winrate
* against regular pachi. Below 1200 is bad (50% winrate and worse), more
* gives diminishing returns (1500 -> 78%, 2000 -> 70% ...) */
p->dcnn_eqex = 1300;
p->joseki_eqex = -200;
p->cfgdn = -1;
/* Even number! */
p->eqex = board_large(b) ? 20 : 14;
p->prune_ladders = true;
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, "eqex") && optval) {
p->eqex = atoi(optval);
/* In the following settings, you can use negative
* numbers to give the hundredths of default eqex.
* E.g. -100 is default eqex, -50 is half of the
* default eqex, -200 is double the default eqex. */
} else if (!strcasecmp(optname, "even") && optval) {
p->even_eqex = atoi(optval);
} else if (!strcasecmp(optname, "policy") && optval) {
p->policy_eqex = atoi(optval);
} else if (!strcasecmp(optname, "b19") && optval) {
p->b19_eqex = atoi(optval);
} else if (!strcasecmp(optname, "cfgd") && optval) {
/* cfgd=3%40%20%20 - 3 levels; immediate libs
* of last move => 40 wins, their neighbors
* 20 wins, 2nd-level neighbors 20 wins;
* neighbors are group-transitive. */
p->cfgdn = atoi(optval); optval += strcspn(optval, "%");
p->cfgd_eqex = calloc2(p->cfgdn + 1, sizeof(*p->cfgd_eqex));
p->cfgd_eqex[0] = 0;
int i;
for (i = 1; *optval; i++, optval += strcspn(optval, "%")) {
optval++;
p->cfgd_eqex[i] = atoi(optval);
}
if (i != p->cfgdn + 1) {
fprintf(stderr, "uct: Missing prior cfdn level %d/%d\n", i, p->cfgdn);
exit(1);
}
} else if (!strcasecmp(optname, "joseki") && optval) {
p->joseki_eqex = atoi(optval);
} else if (!strcasecmp(optname, "eye") && optval) {
p->eye_eqex = atoi(optval);
} else if (!strcasecmp(optname, "ko") && optval) {
p->ko_eqex = atoi(optval);
} else if (!strcasecmp(optname, "pattern") && optval) {
/* Pattern-based prior eqex. */
/* Note that this prior is still going to be
* used only if you have downloaded or
* generated the pattern files! */
p->pattern_eqex = atoi(optval);
} else if (!strcasecmp(optname, "plugin") && optval) {
/* Unlike others, this is just a *recommendation*. */
p->plugin_eqex = atoi(optval);
} else if (!strcasecmp(optname, "prune_ladders")) {
p->prune_ladders = !optval || atoi(optval);
#ifdef DCNN
} else if (!strcasecmp(optname, "dcnn") && optval) {
p->dcnn_eqex = atoi(optval);
#endif
} else {
fprintf(stderr, "uct: Invalid prior argument %s or missing value\n", optname);
exit(1);
}
}
}
if (p->even_eqex < 0) p->even_eqex = p->eqex * -p->even_eqex / 100;
if (p->policy_eqex < 0) p->policy_eqex = p->eqex * -p->policy_eqex / 100;
if (p->b19_eqex < 0) p->b19_eqex = p->eqex * -p->b19_eqex / 100;
if (p->eye_eqex < 0) p->eye_eqex = p->eqex * -p->eye_eqex / 100;
if (p->ko_eqex < 0) p->ko_eqex = p->eqex * -p->ko_eqex / 100;
if (p->joseki_eqex < 0) p->joseki_eqex = p->eqex * -p->joseki_eqex / 100;
if (p->pattern_eqex < 0) p->pattern_eqex = p->eqex * -p->pattern_eqex / 100;
if (p->plugin_eqex < 0) p->plugin_eqex = p->eqex * -p->plugin_eqex / 100;
if (p->dcnn_eqex < 0) p->dcnn_eqex = p->eqex * -p->dcnn_eqex / 100;
if (!using_dcnn(b))
p->dcnn_eqex = 0;
if (p->cfgdn < 0) {
static int large_bonuses[] = { 0, 55, 50, 15 };
static int small_bonuses[] = { 0, 45, 40, 15 };
p->cfgdn = 3;
p->cfgd_eqex = calloc2(p->cfgdn + 1, sizeof(*p->cfgd_eqex));
memcpy(p->cfgd_eqex, board_large(b) ? large_bonuses : small_bonuses, sizeof(large_bonuses));
}
if (p->cfgdn > TREE_NODE_D_MAX) {
fprintf(stderr, "uct: CFG distances only up to %d available\n", TREE_NODE_D_MAX);
exit(1);
}
if (p->pattern_eqex)
u->want_pat = true;
return p;
}
void
uct_prior_done(struct uct_prior *p)
{
assert(p->cfgd_eqex);
free(p->cfgd_eqex);
free(p);
}