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. */
34 floating_t resign_threshold
, sure_win_threshold
;
35 double best2_ratio
, bestr_ratio
;
36 floating_t max_maintime_ratio
;
38 bool territory_scoring
;
40 bool playout_amaf
, playout_amaf_nakade
;
42 int playout_amaf_cutoff
;
47 unsigned long max_tree_size
;
48 unsigned long max_pruned_size
;
49 unsigned long pruning_threshold
;
51 int significant_threshold
;
54 enum uct_thread_model
{
55 TM_TREE
, /* Tree parallelization w/o virtual loss. */
56 TM_TREEVL
, /* Tree parallelization with virtual loss. */
59 bool pondering_opt
; /* User wants pondering */
60 bool pondering
; /* Actually pondering now */
61 bool slave
; /* Act as slave in distributed engine. */
69 struct uct_dynkomi
*dynkomi
;
75 int random_policy_chance
;
78 floating_t local_tree_aging
;
79 #define LTREE_PLAYOUTS_MULTIPLIER 100
80 floating_t local_tree_depth_decay
;
81 bool local_tree_allseq
;
82 bool local_tree_rootseqval
;
86 struct uct_policy
*policy
;
87 struct uct_policy
*random_policy
;
88 struct playout_policy
*playout
;
89 struct uct_prior
*prior
;
90 struct uct_pluginset
*plugins
;
91 struct joseki_dict
*jdict
;
93 /* Used within frame of single genmove. */
94 struct board_ownermap ownermap
;
95 /* Used for coordination among slaves of the distributed engine. */
100 int played_all
; /* games played by all slaves */
102 /* Game state - maintained by setup_state(), reset_state(). */
106 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
108 bool uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
);
110 void uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
);
111 void uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
);
112 void uct_pondering_stop(struct uct
*u
);
115 /* This is the state used for descending the tree; we use this wrapper
116 * structure in order to be able to easily descend in multiple trees
117 * in parallel (e.g. main tree and local tree) or compute cummulative
118 * "path value" throughout the tree descent. */
120 /* Active tree nodes: */
121 struct tree_node
*node
; /* Main tree. */
122 struct tree_node
*lnode
; /* Local tree. */
123 /* Value of main tree node (with all value factors, but unbiased
124 * - without exploration factor), from black's perspective. */
125 struct move_stats value
;
129 typedef struct tree_node
*(*uctp_choose
)(struct uct_policy
*p
, struct tree_node
*node
, struct board
*b
, enum stone color
, coord_t exclude
);
130 typedef floating_t (*uctp_evaluate
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
);
131 typedef void (*uctp_descend
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
);
132 typedef void (*uctp_winner
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
);
133 typedef void (*uctp_prior
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, struct board
*b
, enum stone color
, int parity
);
134 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
);
140 uctp_evaluate evaluate
;
141 uctp_descend descend
;