1 #ifndef PACHI_UCT_INTERNAL_H
2 #define PACHI_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. */
37 floating_t resign_threshold
, sure_win_threshold
;
38 double best2_ratio
, bestr_ratio
;
39 floating_t max_maintime_ratio
;
41 bool territory_scoring
;
43 bool playout_amaf
, playout_amaf_nakade
;
45 int playout_amaf_cutoff
;
50 unsigned long max_tree_size
;
51 unsigned long max_pruned_size
;
52 unsigned long pruning_threshold
;
54 int significant_threshold
;
57 enum uct_thread_model
{
58 TM_TREE
, /* Tree parallelization w/o virtual loss. */
59 TM_TREEVL
, /* Tree parallelization with virtual loss. */
62 bool pondering_opt
; /* User wants pondering */
63 bool pondering
; /* Actually pondering now */
64 bool slave
; /* Act as slave in distributed engine. */
72 struct uct_dynkomi
*dynkomi
;
78 int random_policy_chance
;
81 floating_t local_tree_aging
;
82 #define LTREE_PLAYOUTS_MULTIPLIER 100
83 floating_t local_tree_depth_decay
;
84 bool local_tree_allseq
;
85 bool local_tree_rootseqval
;
89 struct uct_policy
*policy
;
90 struct uct_policy
*random_policy
;
91 struct playout_policy
*playout
;
92 struct uct_prior
*prior
;
93 struct uct_pluginset
*plugins
;
94 struct joseki_dict
*jdict
;
96 /* Used within frame of single genmove. */
97 struct board_ownermap ownermap
;
98 /* Used for coordination among slaves of the distributed engine. */
103 int played_all
; /* games played by all slaves */
105 /* Game state - maintained by setup_state(), reset_state(). */
109 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
111 bool uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
);
113 void uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
);
114 void uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
);
115 void uct_pondering_stop(struct uct
*u
);
118 /* This is the state used for descending the tree; we use this wrapper
119 * structure in order to be able to easily descend in multiple trees
120 * in parallel (e.g. main tree and local tree) or compute cummulative
121 * "path value" throughout the tree descent. */
123 /* Active tree nodes: */
124 struct tree_node
*node
; /* Main tree. */
125 struct tree_node
*lnode
; /* Local tree. */
126 /* Value of main tree node (with all value factors, but unbiased
127 * - without exploration factor), from black's perspective. */
128 struct move_stats value
;
132 typedef struct tree_node
*(*uctp_choose
)(struct uct_policy
*p
, struct tree_node
*node
, struct board
*b
, enum stone color
, coord_t exclude
);
133 typedef floating_t (*uctp_evaluate
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
);
134 typedef void (*uctp_descend
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
);
135 typedef void (*uctp_winner
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
);
136 typedef void (*uctp_prior
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, struct board
*b
, enum stone color
, int parity
);
137 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
, struct board
*final_board
, floating_t result
);
143 uctp_evaluate evaluate
;
144 uctp_descend descend
;