uct/internal.h

   1 #ifndef ZZGO_UCT_INTERNAL_H
   2 #define ZZGO_UCT_INTERNAL_H
   3
   4 #include <signal.h> // sig_atomic_t
   5
   6 #include "debug.h"
   7 #include "move.h"
   8 #include "ownermap.h"
   9 #include "playout.h"
  10 #include "stats.h"
  11 #include "timeinfo.h"
  12 #include "distributed/distributed.h"
  13
  14 struct tree;
  15 struct tree_node;
  16 struct uct_policy;
  17 struct uct_prior;
  18 struct uct_dynkomi;
  19
  20 /* Internal UCT structures */
  21
  22 /* How often to inspect the tree from the main thread to check for playout
  23  * stop, progress reports, etc. (in seconds) */
  24 #define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */
  25
  26 /* Distributed stats for each child of the root node. */
  27 struct node_stats {
  28         struct move_stats2 last_sent_own;
  29         struct move_stats2 added_from_others;
  30         struct tree_node *node;
  31 };
  32
  33 /* Internal engine state. */
  34 struct uct {
  35         int debug_level;
  36         int games, gamelen;
  37         float resign_ratio;
  38         float loss_threshold;
  39         double best2_ratio, bestr_ratio;
  40         bool pass_all_alive;
  41         bool territory_scoring;
  42         int expand_p;
  43         bool playout_amaf, playout_amaf_nakade;
  44         bool amaf_prior;
  45         int playout_amaf_cutoff;
  46         int dumpthres;
  47         int force_seed;
  48         bool no_book;
  49         bool fast_alloc;
  50         unsigned long max_tree_size;
  51         int mercymin;
  52
  53         int threads;
  54         enum uct_thread_model {
  55                 TM_ROOT, /* Root parallelization. */
  56                 TM_TREE, /* Tree parallelization w/o virtual loss. */
  57                 TM_TREEVL, /* Tree parallelization with virtual loss. */
  58         } thread_model;
  59         bool parallel_tree;
  60         bool virtual_loss;
  61         bool pondering_opt; /* User wants pondering */
  62         bool pondering; /* Actually pondering now */
  63         bool slave; /* Act as slave in distributed engine. */
  64         enum stone my_color;
  65
  66         int fuseki_end;
  67         int yose_start;
  68
  69         int dynkomi_mask;
  70         int dynkomi_interval;
  71         struct uct_dynkomi *dynkomi;
  72
  73         float val_scale;
  74         int val_points;
  75         bool val_extra;
  76
  77         int random_policy_chance;
  78         int local_tree;
  79         int tenuki_d;
  80         float local_tree_aging;
  81         bool local_tree_allseq;
  82         /* Playout-localtree integration. */
  83         bool local_tree_playout; // can be true only if ELO playout
  84         bool local_tree_pseqroot;
  85
  86         char *banner;
  87
  88         struct uct_policy *policy;
  89         struct uct_policy *random_policy;
  90         struct playout_policy *playout;
  91         struct uct_prior *prior;
  92
  93         /* Used within frame of single genmove. */
  94         struct board_ownermap ownermap;
  95         /* Used for coordination among slaves of the distributed engine. */
  96         struct node_stats *stats;
  97         int played_own;
  98         int played_all; /* games played by all slaves */
  99
 100         /* Game state - maintained by setup_state(), reset_state(). */
 101         struct tree *t;
 102 };
 103
 104 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
 105
 106 extern volatile sig_atomic_t uct_halt;
 107 extern __thread int thread_id;
 108 extern bool thread_manager_running;
 109
 110 bool uct_pass_is_safe(struct uct *u, struct board *b, enum stone color, bool pass_all_alive);
 111
 112 void uct_prepare_move(struct uct *u, struct board *b, enum stone color);
 113 void uct_genmove_setup(struct uct *u, struct board *b, enum stone color);
 114
 115 /* Progress information of the on-going MCTS search - when did we
 116  * last adjusted dynkomi, printed out stuff, etc. */
 117 struct spawn_ctx;
 118 struct uct_search_state {
 119         /* Number of games simulated for this simulation before
 120          * we started the search. (We have simulated them earlier.) */
 121         int base_playouts;
 122         /* Number of last dynkomi adjustment. */
 123         int last_dynkomi;
 124         /* Number of last game with progress print. */
 125         int last_print;
 126         /* Number of simulations to wait before next print. */
 127         int print_interval;
 128         /* Printed notification about full memory? */
 129         bool print_fullmem;
 130
 131         struct time_stop stop;
 132         struct spawn_ctx *ctx;
 133 };
 134 int uct_search_games(struct uct_search_state *s);
 135 void uct_search_start(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti, struct uct_search_state *s);
 136 void uct_search_progress(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti, struct uct_search_state *s, int i);
 137 bool uct_search_check_stop(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti, struct uct_search_state *s, int i);
 138 struct tree_node *uct_search_best(struct uct *u, struct board *b, enum stone color, bool pass_all_alive, int played_games, int base_playouts, coord_t *best_coord);
 139
 140
 141 /* This is the state used for descending the tree; we use this wrapper
 142  * structure in order to be able to easily descend in multiple trees
 143  * in parallel (e.g. main tree and local tree) or compute cummulative
 144  * "path value" throughout the tree descent. */
 145 struct uct_descent {
 146         /* Active tree nodes: */
 147         struct tree_node *node; /* Main tree. */
 148         struct tree_node *lnode; /* Local tree. */
 149         /* Value of main tree node (with all value factors, but unbiased
 150          * - without exploration factor), from black's perspective. */
 151         struct move_stats value;
 152 };
 153
 154
 155 typedef struct tree_node *(*uctp_choose)(struct uct_policy *p, struct tree_node *node, struct board *b, enum stone color, coord_t exclude);
 156 typedef float (*uctp_evaluate)(struct uct_policy *p, struct tree *tree, struct uct_descent *descent, int parity);
 157 typedef void (*uctp_descend)(struct uct_policy *p, struct tree *tree, struct uct_descent *descent, int parity, bool allow_pass);
 158 typedef void (*uctp_winner)(struct uct_policy *p, struct tree *tree, struct uct_descent *descent);
 159 typedef void (*uctp_prior)(struct uct_policy *p, struct tree *tree, struct tree_node *node, struct board *b, enum stone color, int parity);
 160 typedef void (*uctp_update)(struct uct_policy *p, struct tree *tree, struct tree_node *node, enum stone node_color, enum stone player_color, struct playout_amafmap *amaf, float result);
 161
 162 struct uct_policy {
 163         struct uct *uct;
 164         uctp_choose choose;
 165         uctp_winner winner;
 166         uctp_evaluate evaluate;
 167         uctp_descend descend;
 168         uctp_update update;
 169         uctp_prior prior;
 170         bool wants_amaf;
 171         void *data;
 172 };
 173
 174 #endif