1 #ifndef ZZGO_UCT_INTERNAL_H
2 #define ZZGO_UCT_INTERNAL_H
4 /* Internal UCT structures */
20 /* How big proportion of ownermap counts must be of one color to consider
23 /* How many games to consider at minimum before judging groups. */
24 #define GJ_MINGAMES 500
26 /* Internal engine state. */
30 float resign_threshold
, sure_win_threshold
;
31 double best2_ratio
, bestr_ratio
;
33 bool territory_scoring
;
35 bool playout_amaf
, playout_amaf_nakade
;
37 int playout_amaf_cutoff
;
42 unsigned long max_tree_size
;
43 unsigned long max_pruned_size
;
44 unsigned long pruning_threshold
;
46 int significant_threshold
;
49 enum uct_thread_model
{
50 TM_TREE
, /* Tree parallelization w/o virtual loss. */
51 TM_TREEVL
, /* Tree parallelization with virtual loss. */
54 bool pondering_opt
; /* User wants pondering */
55 bool pondering
; /* Actually pondering now */
56 bool slave
; /* Act as slave in distributed engine. */
64 struct uct_dynkomi
*dynkomi
;
70 int treepool_chance
[2];
79 int treepool_pickfactor
;
81 int random_policy_chance
;
84 float local_tree_aging
;
85 bool local_tree_allseq
;
86 /* Playout-localtree integration. */
87 bool local_tree_playout
; // can be true only if ELO playout
88 bool local_tree_pseqroot
;
92 struct uct_policy
*policy
;
93 struct uct_policy
*random_policy
;
94 struct playout_policy
*playout
;
95 struct uct_prior
*prior
;
96 struct uct_pluginset
*plugins
;
97 struct joseki_dict
*jdict
;
99 /* Used within frame of single genmove. */
100 struct board_ownermap ownermap
;
101 /* Used for coordination among slaves of the distributed engine. */
106 int played_all
; /* games played by all slaves */
108 /* Game state - maintained by setup_state(), reset_state(). */
112 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
114 bool uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
);
116 void uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
);
117 void uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
);
118 void uct_pondering_stop(struct uct
*u
);
121 /* This is the state used for descending the tree; we use this wrapper
122 * structure in order to be able to easily descend in multiple trees
123 * in parallel (e.g. main tree and local tree) or compute cummulative
124 * "path value" throughout the tree descent. */
126 /* Active tree nodes: */
127 struct tree_node
*node
; /* Main tree. */
128 struct tree_node
*lnode
; /* Local tree. */
129 /* Value of main tree node (with all value factors, but unbiased
130 * - without exploration factor), from black's perspective. */
131 struct move_stats value
;
135 typedef struct tree_node
*(*uctp_choose
)(struct uct_policy
*p
, struct tree_node
*node
, struct board
*b
, enum stone color
, coord_t exclude
);
136 typedef float (*uctp_evaluate
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
);
137 typedef void (*uctp_descend
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
);
138 typedef void (*uctp_winner
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
);
139 typedef void (*uctp_prior
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, struct board
*b
, enum stone color
, int parity
);
140 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
);
146 uctp_evaluate evaluate
;
147 uctp_descend descend
;