1 #ifndef PACHI_UCT_INTERNAL_H
2 #define PACHI_UCT_INTERNAL_H
4 /* Internal UCT structures */
10 #include "patternsp.h"
11 #include "patternprob.h"
23 /* How big proportion of ownermap counts must be of one color to consider
26 /* How many games to consider at minimum before judging groups. */
27 #define GJ_MINGAMES 500
29 /* Internal engine state. */
40 floating_t resign_threshold
, sure_win_threshold
;
41 double best2_ratio
, bestr_ratio
;
42 floating_t max_maintime_ratio
;
43 bool pass_all_alive
; /* Current value */
44 bool allow_losing_pass
;
45 bool territory_scoring
;
49 int playout_amaf_cutoff
;
54 unsigned long max_tree_size
;
55 unsigned long max_pruned_size
;
56 unsigned long pruning_threshold
;
58 int significant_threshold
;
61 enum uct_thread_model
{
62 TM_TREE
, /* Tree parallelization w/o virtual loss. */
63 TM_TREEVL
, /* Tree parallelization with virtual loss. */
66 bool pondering_opt
; /* User wants pondering */
67 bool pondering
; /* Actually pondering now */
68 bool slave
; /* Act as slave in distributed engine. */
69 int max_slaves
; /* Optional, -1 if not set */
70 int slave_index
; /* 0..max_slaves-1, or -1 if not set */
78 struct uct_dynkomi
*dynkomi
;
85 floating_t val_bytemp_min
;
87 int random_policy_chance
;
90 floating_t local_tree_aging
;
91 #define LTREE_PLAYOUTS_MULTIPLIER 100
92 floating_t local_tree_depth_decay
;
93 bool local_tree_allseq
;
94 bool local_tree_neival
;
100 bool local_tree_rootchoose
;
104 struct uct_policy
*policy
;
105 struct uct_policy
*random_policy
;
106 struct playout_policy
*playout
;
107 struct uct_prior
*prior
;
108 struct uct_pluginset
*plugins
;
109 struct joseki_dict
*jdict
;
111 struct pattern_setup pat
;
112 /* Various modules (prior, policy, ...) set this if they want pattern
113 * database to be loaded. */
116 /* Used within frame of single genmove. */
117 struct board_ownermap ownermap
;
118 /* Used for coordination among slaves of the distributed engine. */
122 double stats_delay
; /* stored in seconds */
124 int played_all
; /* games played by all slaves */
126 /* Game state - maintained by setup_state(), reset_state(). */
130 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
132 bool uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
);
134 void uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
);
135 void uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
);
136 void uct_pondering_stop(struct uct
*u
);
139 /* This is the state used for descending the tree; we use this wrapper
140 * structure in order to be able to easily descend in multiple trees
141 * in parallel (e.g. main tree and local tree) or compute cummulative
142 * "path value" throughout the tree descent. */
144 /* Active tree nodes: */
145 struct tree_node
*node
; /* Main tree. */
146 struct tree_node
*lnode
; /* Local tree. */
147 /* Value of main tree node (with all value factors, but unbiased
148 * - without exploration factor), from black's perspective. */
149 struct move_stats value
;
153 typedef struct tree_node
*(*uctp_choose
)(struct uct_policy
*p
, struct tree_node
*node
, struct board
*b
, enum stone color
, coord_t exclude
);
154 typedef floating_t (*uctp_evaluate
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
);
155 typedef void (*uctp_descend
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
);
156 typedef void (*uctp_winner
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
);
157 typedef void (*uctp_prior
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, struct board
*b
, enum stone color
, int parity
);
158 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
);
164 uctp_evaluate evaluate
;
165 uctp_descend descend
;