playout_elo_choose(): Pass when no moves remain in the probdist
[pachi.git] / playout / elo.c
blob92f76e8ac49982a806456f9825910f9846c3706a
1 /* Playout player based on probability distribution generated over
2 * the available moves. */
4 /* We use the ELO-based (Coulom, 2007) approach, where each board
5 * feature (matched pattern, self-atari, capture, MC owner?, ...)
6 * is pre-assigned "playing strength" (gamma).
8 * Then, the problem of choosing a move is basically a team
9 * competition in ELO terms - each spot is represented by a team
10 * of features appearing there; the team gamma is product of feature
11 * gammas. The team gammas make for a probability distribution of
12 * moves to be played.
14 * We use the general pattern classifier that will find the features
15 * for us, and external datasets that can be harvested from a set
16 * of game records (see the HACKING file for details): patterns.spat
17 * as a dictionary of spatial stone configurations, and patterns.gamma
18 * with strengths of particular features. */
20 #include <assert.h>
21 #include <math.h>
22 #include <stdio.h>
23 #include <stdlib.h>
25 #define DEBUG
26 #include "board.h"
27 #include "debug.h"
28 #include "pattern.h"
29 #include "patternsp.h"
30 #include "playout.h"
31 #include "playout/elo.h"
32 #include "random.h"
33 #include "tactics.h"
34 #include "uct/prior.h"
36 #define PLDEBUGL(n) DEBUGL_(p->debug_level, n)
39 /* Note that the context can be shared by multiple threads! */
41 struct patternset {
42 pattern_spec ps;
43 struct pattern_config pc;
44 struct features_gamma *fg;
47 struct elo_policy {
48 float selfatari;
49 struct patternset choose, assess;
53 /* This is the core of the policy - initializes and constructs the
54 * probability distribution over the move candidates. */
56 int
57 elo_get_probdist(struct playout_policy *p, struct patternset *ps, struct board *b, enum stone to_play, struct probdist *pd)
59 //struct elo_policy *pp = p->data;
60 int moves = 0;
62 /* First, assign per-point probabilities. */
64 for (int f = 0; f < b->flen; f++) {
65 struct move m = { .coord = b->f[f], .color = to_play };
67 /* Skip pass (for now)? */
68 if (is_pass(m.coord)) {
69 skip_move:
70 probdist_set(pd, f, 0);
71 continue;
73 //fprintf(stderr, "<%d> %s\n", f, coord2sstr(m.coord, b));
75 /* Skip invalid moves. */
76 if (!board_is_valid_move(b, &m))
77 goto skip_move;
79 /* We shall never fill our own single-point eyes. */
80 /* XXX: In some rare situations, this prunes the best move:
81 * Bulk-five nakade with eye at 1-1 point. */
82 if (board_is_one_point_eye(b, &m.coord, to_play)) {
83 goto skip_move;
86 moves++;
87 /* Each valid move starts with gamma 1. */
88 float g = 1.f;
90 /* Some easy features: */
91 /* XXX: We just disable them for now since we call the
92 * pattern matcher; you need the gammas file. */
93 #if 0
94 if (is_bad_selfatari(b, to_play, m.coord))
95 g *= pp->selfatari;
96 #endif
98 /* Match pattern features: */
99 struct pattern p;
100 pattern_match(&ps->pc, ps->ps, &p, b, &m);
101 for (int i = 0; i < p.n; i++) {
102 /* Multiply together gammas of all pattern features. */
103 float gamma = feature_gamma(ps->fg, &p.f[i], NULL);
104 //char buf[256] = ""; feature2str(buf, &p.f[i]);
105 //fprintf(stderr, "<%d> %s feat %s gamma %f\n", f, coord2sstr(m.coord, b), buf, gamma);
106 g *= gamma;
109 probdist_set(pd, f, g);
110 //fprintf(stderr, "<%d> %s %f (E %f)\n", f, coord2sstr(m.coord, b), probdist_one(pd, f), pd->items[f]);
113 return moves;
117 coord_t
118 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
120 #ifdef BOARD_GAMMA
121 struct probdist *pd = &b->prob[to_play - 1];
122 /* Make sure ko-prohibited move does not get picked. */
123 if (!is_pass(b->ko.coord)) {
124 assert(b->ko.color == to_play);
125 probdist_set(pd, b->ko.coord, 0);
127 /* TODO: FEAT_CONTIGUITY support. */
128 /* Pick a move. */
129 coord_t c = pd->total >= 0.01 ? probdist_pick(pd) : pass;
130 /* Repair the damage. */
131 if (!is_pass(b->ko.coord))
132 board_gamma_update(b, b->ko.coord, to_play);
133 return c;
134 #else
135 struct elo_policy *pp = p->data;
136 float pdi[b->flen]; memset(pdi, 0, sizeof(pdi));
137 struct probdist pd = { .n = b->flen, .items = pdi, .total = 0 };
138 elo_get_probdist(p, &pp->choose, b, to_play, &pd);
139 int f = probdist_pick(&pd);
140 return b->f[f];
141 #endif
144 void
145 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
147 struct elo_policy *pp = p->data;
148 float pdi[map->b->flen]; memset(pdi, 0, sizeof(pdi));
149 struct probdist pd = { .n = map->b->flen, .items = pdi, .total = 0 };
151 int moves;
152 moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
154 /* It is a question how to transform the gamma to won games; we use
155 * a naive approach currently, but not sure how well it works. */
156 /* TODO: Try sqrt(p), atan(p)/pi*2. */
158 for (int f = 0; f < map->b->flen; f++) {
159 coord_t c = map->b->f[f];
160 if (!map->consider[c])
161 continue;
162 add_prior_value(map, c, probdist_one(&pd, f) / probdist_total(&pd), games);
166 void
167 playout_elo_done(struct playout_policy *p)
169 struct elo_policy *pp = p->data;
170 features_gamma_done(pp->choose.fg);
171 features_gamma_done(pp->assess.fg);
175 struct playout_policy *
176 playout_elo_init(char *arg, struct board *b)
178 struct playout_policy *p = calloc(1, sizeof(*p));
179 struct elo_policy *pp = calloc(1, sizeof(*pp));
180 p->data = pp;
181 p->choose = playout_elo_choose;
182 p->assess = playout_elo_assess;
183 p->done = playout_elo_done;
185 const char *gammafile = features_gamma_filename;
186 /* Some defaults based on the table in Remi Coulom's paper. */
187 pp->selfatari = 0.06;
189 struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
190 int xspat = -1;
192 if (arg) {
193 char *optspec, *next = arg;
194 while (*next) {
195 optspec = next;
196 next += strcspn(next, ":");
197 if (*next) { *next++ = 0; } else { *next = 0; }
199 char *optname = optspec;
200 char *optval = strchr(optspec, '=');
201 if (optval) *optval++ = 0;
203 if (!strcasecmp(optname, "selfatari") && optval) {
204 pp->selfatari = atof(optval);
205 } else if (!strcasecmp(optname, "gammafile") && optval) {
206 /* patterns.gamma by default. We use this,
207 * and need also ${gammafile}f (e.g.
208 * patterns.gammaf) for fast (MC) features. */
209 gammafile = strdup(optval);
210 } else if (!strcasecmp(optname, "xspat") && optval) {
211 /* xspat==0: don't match spatial features
212 * xspat==1: match *only* spatial features */
213 xspat = atoi(optval);
214 } else {
215 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
216 exit(1);
221 pc.spat_dict = spatial_dict_init(false);
223 pp->assess.pc = pc;
224 pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
225 memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
226 for (int i = 0; i < FEAT_MAX; i++)
227 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
228 pp->assess.ps[i] = 0;
230 /* In playouts, we need to operate with much smaller set of features
231 * in order to keep reasonable speed. */
232 /* TODO: Configurable. */ /* TODO: Tune. */
233 pp->choose.pc = FAST_PATTERN_CONFIG;
234 pp->choose.pc.spat_dict = pc.spat_dict;
235 char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
236 pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
237 memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
238 for (int i = 0; i < FEAT_MAX; i++)
239 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
240 pp->choose.ps[i] = 0;
242 #ifdef BOARD_GAMMA
243 b->gamma = pp->choose.fg;
244 #endif
246 return p;