/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
* This is GNU Go, a Go program. Contact gnugo@gnu.org, or see *
* http://www.gnu.org/software/gnugo/ for more information. *
* Copyright 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, *
* 2008 and 2009 by the Free Software Foundation. *
* This program is free software; you can redistribute it and/or *
* modify it under the terms of the GNU General Public License as *
* published by the Free Software Foundation - version 3 or *
* (at your option) any later version. *
* This program is distributed in the hope that it will be useful, *
* but WITHOUT ANY WARRANTY; without even the implied warranty of *
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the *
* GNU General Public License in file COPYING for more details. *
* You should have received a copy of the GNU General Public *
* License along with this program; if not, write to the Free *
* Software Foundation, Inc., 51 Franklin Street, Fifth Floor, *
* Boston, MA 02111, USA. *
\* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
/* Extract fuseki patterns from the initial moves of a collection
* This program finds the most common positions from the initial moves
* of a collection of games, and generates patterns in patterns.db
* format for the most common moves in these positions.
* Positions are identified by Zobrist hash values, completely
* ignoring the risk for hash collisions. In order to take all
* symmetries into account, we compute 8 hash values, one for each
* transformation of the board. Rather than playing on 8 boards in
* parallel, we construct 8 transformed copies of the Zobrist hash
* tables and compute one hash value for each of these. To get a
* transformation invariant hash, we finally sort the 8 hash values.
* extract_fuseki sgflist 9 8 400
* generates (up to) 400 patterns, considering the 8 first moves of
* the 9x9 games listed in the file sgflist, and writes the patterns
* to stdout. sgflist is a file containing sgf filenames, one per line.
* The generated patterns may look like, e.g.
* The comment line gives the information that this position has been
* found 18 times among the analyzed games, and 3 out of these 18 times,
* the move * has been played. The same number 3 is entered as pattern
* value on the colon line for use by the fuseki module.
* Notes on the statistics:
* The statistics code assumes that every input file is valid. Use
* the output file option to sort out which input files are valid, and
* check output for problems. Rerun after fixing/removing invalid files.
* Outcome is defined by RE in sgf. Files without a parsable RE, or which
* do not have a winner, are invalid and need to be excluded.
* Pearson chi squared at 5% is used to test significance of
* differences among moves at a given position. Moves excluded by
* popularity rules are grouped together and considered as one. A
* positive result means that among all possible moves in a position,
* there's a difference somewhere. The next question is where. One
* clue comes from dchisq, which is the contribution to the overall
* chi squared for each move, with larger meaning higher impact on
* significance of overall result. Another comes from post hoc tests.
* Each pair of moves from a position with a statistically significant
* impact of move choice is compared, again with Pearson chi squared
* at 5%, and the positive tests printed. No correction is done for
* multiple tests. Pairs with less than 6 total moves are not tested,
* so it's possible for there to be a significant overall result
* without any positive post hocs. In this case, the overall result is
* If the interest is solely in statistics, using min_pos_freq to
* avoid positions without enough data to hope for significance makes
* Note that the popularity exclusion rules can result in patterns being
* left in the db which have no parent in the db.
#include "../sgf/sgftree.h"
Usage: extract_fuseki files boardsize moves patterns handicap strength half_board min_pos_freq min_move_percent min_move_freq [output file]\n\
files: The name of a file listing sgf files to examine,\n\
one filename per line.\n\
boardsize: Only consider games with this size.\n\
moves: Number of moves considered in each game.\n\
handicap: 0 - no handicap, 1 - any game, 2-9 - two to nine handicap stones\n\
strength: The lowest strength of the players (1k-30k)\n\
half_board: 0 - full board patterns, 1 - half board patterns\n\
min_pos_freq: how many times a position must occur before patterns\n\
min_move_percent: minimum popularity relative to most popular move \n\
(counted by unique players) required of a move \n\
in a given position before it gets a pattern\n\
min_move_freq: minimum number of unique players who must play a move\n\
before it gets a pattern\n\
output file: Optional (if this exists, extract_fuseki will sort the games instead)\n\
/* Maximum length of sgf filename. */
/* Number of moves to consider in each game, given as argument.*/
/* Flag checking the setting for generating half board patterns */
int half_board_patterns
= 0;
/* Maximum number of patterns to generate */
#define MAX_PATTERNS_TO_EXTRACT 100000
/* Handicap value, given as argument.*/
/* Lowest strength, given as argument.*/
/* Min # of times a position must be seen before moves from it become
* Might want this larger to ensure reasonable statistics, 6 or more, say
* or smaller to hit every move down to unique games, 2 say;
* or even keep churning out moves with 1.
/* popularity arguments */
/* Number of games to analyze. */
/* Dynamically allocated array marking the games that could not be
/* WARN 1 warns about unused games. */
/* WARN 2 also notes assumptions about metainfo. */
/* Dynamically allocated list of sgf file names. */
/* Zobrist hash tables, rotated and reflected into all 8 transformations. */
unsigned int O_hash
[8][MAX_BOARD
][MAX_BOARD
];
unsigned int X_hash
[8][MAX_BOARD
][MAX_BOARD
];
unsigned int move_hash
[8][MAX_BOARD
][MAX_BOARD
];
/* A board is hashed 8 times, once for each transformation, and these
* numbers are sorted into a transformation invariant hash.
/* This is defined in engine/matchpat.c */
extern const int transformations
[8][2][2];
/* A situation is the combination of a board position and the move to
* be made. We use the invariant hashes excluding and including the move
* as identification. If are interested in positions, we only use the first
* We ignore the possibility of a hash collision.
* outcome is the color which won the game
* player is the (hashed) name of the player who made the move
struct invariant_hash pre
;
struct invariant_hash post
;
/* Dynamically allocated table of situations encountered in the analysis. */
struct situation
*situation_table
;
int number_of_situations
;
/* Data type for frequencies of e.g. situations or positions. 'index'
* is the index into situation_table.
/* Position frequency table. */
struct frequency
*frequency_table
;
int number_of_distinct_positions
;
/* The most common situations are called winners. These are the ones
* we generate patterns for.
* 'index' is normally an index into situation_table, or -1 for
* special aggregate entry (with no pattern) to collect stats for
* pre is hash[0], and must be stored here for aggregate
char pattern
[MAX_BOARD
][MAX_BOARD
];
/* Dynamically allocated table of winners. */
struct winner
*winning_moves
;
int number_of_winning_moves
;
/* critical values of chisquare distribution with n degrees of freedom */
double chisquarecrit05
[] = {
3.8415, 5.9915, 7.8147, 9.4877, 11.0705, 12.5916, 14.0671, 15.5073,
16.9190, 18.3070, 19.6751, 21.0261, 22.3620, 23.6848, 24.9958, 26.2962,
27.5871, 28.8693, 30.1435, 31.4104, 32.67057, 33.92444, 35.17246,
36.41503, 37.65248, 38.88514, 40.11327, 41.33714, 42.55697, 43.77297,
44.98534, 46.19426, 47.39988, 48.60237, 49.80185, 50.99846, 52.19232,
53.38354, 54.57223, 55.75848, 56.94239, 58.12404, 59.30351, 60.48089,
61.65623, 62.82962, 64.00111, 65.17077, 66.33865, 67.50481};
/* p < 0.10, should be same size as 05 */
double chisquarecrit10
[] = {
2.7055, 4.6052, 6.2514, 7.7794, 9.2364, 10.6446, 12.0170, 13.3616,
14.6837, 15.9872, 17.2750, 18.5493, 19.8119, 21.0641, 22.3071, 23.5418,
24.7690, 25.9894, 27.2036, 28.4120, 29.61509, 30.81328, 32.00690,
33.19624, 34.38159, 35.56317, 36.74122, 37.91592, 39.08747, 40.25602,
41.42174, 42.58475, 43.74518, 44.90316, 46.05879, 47.21217, 48.36341,
49.51258, 50.65977, 51.80506, 52.94851, 54.09020, 55.23019, 56.36854,
57.50530, 58.64054, 59.77429, 60.90661, 62.03754, 63.16712};
double chisquarecrit01
[] = {
6.63489660102121, 9.21034037197618, 11.3448667301444, 13.2767041359876,
15.086272469389, 16.8118938297709, 18.4753069065824, 20.0902350296632,
21.6659943334619, 23.2092511589544, 24.7249703113183, 26.2169673055359,
27.6882496104570, 29.1412377406728, 30.5779141668925, 31.9999269088152,
33.4086636050046, 34.8053057347051, 36.1908691292701, 37.5662347866250,
38.9321726835161, 40.2893604375938, 41.6383981188585, 42.9798201393516,
44.3141048962192, 45.6416826662832, 46.9629421247514, 48.2782357703155,
49.5878844728988, 50.8921813115171, 52.1913948331919, 53.4857718362354,
54.7755397601104, 56.0609087477891, 57.3420734338592, 58.619214501687,
59.8925000450869, 61.1620867636897, 62.4281210161849, 63.6907397515645,
64.9500713352112, 66.2062362839932, 67.4593479223258, 68.7095129693454,
69.9568320658382, 71.2014002483115, 72.4433073765482, 73.6826385201058,
74.9194743084782, 76.1538912490127};
double chisquarecrit001
[] = {
10.8275661706627, 13.8155105579643, 16.2662361962381, 18.4668269529032,
20.5150056524329, 22.4577444848253, 24.3218863478569, 26.1244815583761,
27.8771648712566, 29.5882984450744, 31.26413362024, 32.9094904073602,
34.5281789748709, 36.1232736803981, 37.6972982183538, 39.2523547907685,
40.7902167069025, 42.31239633168, 43.8201959645175, 45.3147466181259,
46.7970380415613, 48.2679422908352, 49.7282324664315, 51.1785977773774,
52.6196557761728, 54.0519623885766, 55.4760202057452, 56.8922853933536,
58.3011734897949, 59.7030643044299, 61.0983060810581, 62.4872190570885,
63.870098522345, 65.2472174609424, 66.618828843701, 67.9851676260242,
69.3464524962412, 70.702887411505, 72.0546629519878, 73.401957518991,
74.7449383984238, 76.0837627077, 77.418578241314, 78.749524228043,
80.076732010819, 81.40032565871, 82.720422519124, 84.0371337172235,
85.350564608593, 86.6608151904032};
* Append the files that are sorted to a specific file
write_sgf_filenames(const char *name
, char *filenames
[])
FILE *namefile
= fopen(name
, "a");
fprintf(stderr
, "Fatal error, couldn't open %s.\n", name
);
for (n
= 0; n
< number_of_games
; n
++) {
if (unused_games
[n
] == 0)
fprintf(namefile
, "%s\n", filenames
[n
]);
/* Read the sgf file names. These are assumed to be stored one per
* line in the file with the name given by 'name'. The sgf file names
* are copied into dynamically allocated memory by strdup() and
* pointers to the names are stored into the 'filenames[]' array. It
* is assumed that 'filenames' has been allocated sufficiently large
* before this this function is called. If 'filenames' is NULL, the
* sgf file names are only counted. The number of sgf file names is
read_sgf_filenames(const char *name
, char *filenames
[])
FILE *namefile
= fopen(name
, "r");
fprintf(stderr
, "Fatal error, couldn't open %s.\n", name
);
while (fgets(buf
, BUFSIZE
, namefile
) != NULL
) {
if (buf
[strlen(buf
) - 2] == '\r') {
buf
[strlen(buf
) - 2] = '\0';
/* Delete carriage return character, if any. */
buf
[strlen(buf
) - 1] = '\0';
/* Delete newline character. */
filenames
[n
] = strdup(buf
);
if (filenames
[n
] == NULL
) {
fprintf(stderr
, "Fatal error, strdup() failed.\n");
/* Fill one of the zobrist hash tables with random numbers. */
init_zobrist_table(unsigned int hash
[8][MAX_BOARD
][MAX_BOARD
])
for (m
= 0; m
< board_size
; m
++)
for (n
= 0; n
< board_size
; n
++) {
for (k
= 0; 32*k
< CHAR_BIT
*sizeof(hash
[0][0][0]); k
++)
hash
[0][m
][n
] |= gg_urand() << k
*32;
/* Fill in all transformations of the hash table. */
for (m
= 0; m
< board_size
; m
++)
for (n
= 0; n
< board_size
; n
++) {
TRANSFORM2(m
-mid
, n
-mid
, &i
, &j
, k
);
hash
[k
][m
][n
] = hash
[0][i
+mid
][j
+mid
];
for (k
= 0; k
< 8; k
++) {
for (m
= 0; m
< board_size
; m
++) {
for (n
= 0; n
< board_size
; n
++)
fprintf(stderr
, "%8x ", hash
[k
][m
][n
]);
/* Initialize all Zobrist hash tables with random numbers. */
init_zobrist_numbers(void)
init_zobrist_table(O_hash
);
init_zobrist_table(X_hash
);
init_zobrist_table(move_hash
);
/* Initialize the situation_table array. */
situation_table
= calloc(moves_per_game
* number_of_games
,
sizeof(*situation_table
));
fprintf(stderr
, "Fatal error, failed to allocate situations table.\n");
number_of_situations
= 0;
/* Compare two hash values. Used for sorting the hash values in the
compare_numbers(const void *a
, const void *b
)
unsigned int aa
= *((const unsigned int *) a
);
unsigned int bb
= *((const unsigned int *) b
);
/* Compute hash values for all transformations of the position
* currently in the p[][] array. The hash values are not sorted by
common_hash_board(struct invariant_hash
*hash
, int color_to_play
)
for (m
= 0; m
< board_size
; m
++)
for (n
= 0; n
< board_size
; n
++) {
for (k
= 0; k
< 8; k
++) {
if (BOARD(m
, n
) == color_to_play
)
hash
->values
[k
] ^= O_hash
[k
][m
][n
];
else if (BOARD(m
, n
) != EMPTY
)
hash
->values
[k
] ^= X_hash
[k
][m
][n
];
/* Compute invariant hash for the current position. */
hash_board(struct invariant_hash
*hash
, int color_to_play
)
common_hash_board(hash
, color_to_play
);
/* Sort the 8 hash values. */
gg_sort(hash
->values
, 8, sizeof(hash
->values
[0]), compare_numbers
);
/* Compute invariant hash for the current situation, i.e. position
* plus a move to be made.
hash_board_and_move(struct invariant_hash
*hash
, int color_to_play
,
common_hash_board(hash
, color_to_play
);
hash
->values
[k
] ^= move_hash
[k
][m
][n
];
/* Notice that we of course must wait with sorting until we have
* added the move to the hash values.
gg_sort(hash
->values
, 8, sizeof(hash
->values
[0]), compare_numbers
);
/* the so called X31 hash from gtk, for hashing strings */
hash_string(const char *v
)
/* Adapted from play_sgf_tree() in engine/sgfutils.c. Returns the
* next move from the game record in (*m, *n) and color in *color. If
* handicap stones are encountered, these are put on the board
* immediately. Return value is 1 if another move was found in the
* game record, 0 otherwise.
get_move_from_sgf(SGFNode
*node
, int *m
, int *n
, int *color
)
for (prop
= node
->props
; prop
; prop
= prop
->next
) {
if (!prop
|| !prop
->name
|| !node
) {
/* something wrong with the SGF file properties */
fprintf(stderr
, "Something wrong with the SGF file properties.\n");
get_moveXY(prop
, &i
, &j
, board_size
);
/* Put handicap stones on the board at once. */
add_stone(POS(i
, j
), BLACK
);
fprintf(stderr
, "Warning: white stone added.\n");
fprintf(stderr
, "Warning: PL property encountered.\n");
*color
= (prop
->name
== SGFW
) ? WHITE
: BLACK
;
if (!get_moveXY(prop
, m
, n
, board_size
)) {
fprintf(stderr
, "Warning: failed to get move coordinates.\n");
/* Add a situation to the situation_table array. */
add_situation(struct invariant_hash
*pre
, struct invariant_hash
*post
,
int outcome
, unsigned int player
)
situation_table
[number_of_situations
].pre
= *pre
;
situation_table
[number_of_situations
].post
= *post
;
situation_table
[number_of_situations
].outcome
= outcome
;
situation_table
[number_of_situations
].player
= player
;
/* Compare two situations. Used (indirectly, see compare_situations2)
* for sorting the situation_table array
* and when building frequency tables for the different moves at the
compare_situations(const void *a
, const void *b
)
const struct situation
*aa
= a
;
const struct situation
*bb
= b
;
for (k
= 0; k
< 8; k
++) {
if (aa
->pre
.values
[k
] > bb
->pre
.values
[k
])
if (aa
->pre
.values
[k
] < bb
->pre
.values
[k
])
for (k
= 0; k
< 8; k
++) {
if (aa
->post
.values
[k
] > bb
->post
.values
[k
])
if (aa
->post
.values
[k
] < bb
->post
.values
[k
])
compare_situations2(const void *a
, const void *b
)
const struct situation
*aa
= a
;
const struct situation
*bb
= b
;
int r
= compare_situations(a
, b
);
if (aa
->player
> bb
->player
)
if (aa
->player
< bb
->player
)
/* Compare two positions. Used when building frequency table for the
* different positions encountered.
compare_positions(const void *a
, const void *b
)
const struct situation
*aa
= a
;
const struct situation
*bb
= b
;
for (k
= 0; k
< 8; k
++) {
if (aa
->pre
.values
[k
] > bb
->pre
.values
[k
])
if (aa
->pre
.values
[k
] < bb
->pre
.values
[k
])
/* Compare two frequency table entries. The returned values are
* "backwards" because we always want to sort frequencies in falling
* The first version counts every game equally, the second version
* counts a game only once per unique player.
compare_frequencies(const void *a
, const void *b
)
const struct frequency
*aa
= a
;
const struct frequency
*bb
= b
;
compare_frequencies2(const void *a
, const void *b
)
const struct frequency
*aa
= a
;
const struct frequency
*bb
= b
;
if (aa
->n_player
< bb
->n_player
)
if (aa
->n_player
> bb
->n_player
)
* find_region answers in what region the move is.
* There are 9 regions, corners, sides and center.
find_region(int m
, int n
)
else if (n
> 6 && m
> 13)
/* otherwise in center */
/* If this situation is listed among the winners, fill in the pattern
* entry of the winner struct.
store_pattern_if_winner(struct invariant_hash
*pre
,
struct invariant_hash
*post
,
for (k
= 0; k
< number_of_winning_moves
; k
++) {
if (winning_moves
[k
].index
!= -1
&& compare_situations(&situation_table
[winning_moves
[k
].index
],
/* This is a winner. Record the pattern. */
for (i
= 0; i
< board_size
; i
++)
for (j
= 0; j
< board_size
; j
++) {
if (BOARD(i
, j
) == EMPTY
)
winning_moves
[k
].pattern
[i
][j
] = '.';
else if (BOARD(i
, j
) == color
) {
winning_moves
[k
].pattern
[i
][j
] = 'O';
else if ((color
== WHITE
&& BOARD(i
, j
) == BLACK
)
|| (color
== BLACK
&& BOARD(i
, j
) == WHITE
)) {
winning_moves
[k
].pattern
[i
][j
] = 'X';
else { /* something is wrong */
fprintf(stderr
, "Error in store_pattern_if_winner: %d\n", k
);
winning_moves
[k
].pattern
[i
][j
] = '.';
winning_moves
[k
].pattern
[m
][n
] = '*';
/* Add ? in areas far away from the move. */
if (half_board_patterns
== 1 && move_number
> 3 && board_size
== 19)
region
= find_region(m
, n
);
for (i
= 0; i
< board_size
; i
++) {
for (j
= 0; j
< board_size
; j
++) {
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
if (i
+ board_size
- j
< 14)
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
winning_moves
[k
].pattern
[i
][j
] = '?';
/* Play through the initial moves of a game. If 'collect_statistics'
* is set, store all encountered situations in the situation_table
* array. 'collect_statistics' will be set to the color which won the
* game. Otherwise, see if there are any winners among the situations
* and store the corresponding pattern so that it can subsequently be
* printed. Return 0 if there was some problem with the game record,
* e.g. fewer moves than expected.
examine_game(SGFNode
*sgf
, int collect_statistics
)
struct invariant_hash prehash
;
struct invariant_hash posthash
;
unsigned int white_player
, black_player
;
if (!sgfGetCharProperty(sgf
, "PW", &PW
))
white_player
= hash_string("");
white_player
= hash_string(PW
);
if (!sgfGetCharProperty(sgf
, "PB", &PB
))
black_player
= hash_string("");
black_player
= hash_string(PB
);
/* Call the engine to clear the board. */
/* Loop through the first moves_per_game moves of each game. */
for (k
= 0; k
< moves_per_game
&& node
!= NULL
; node
= node
->child
) {
if (!get_move_from_sgf(node
, &m
, &n
, &color
)) {
/* something is wrong with the file */
fprintf(stderr
, "move number:%d\n", k
);
gg_assert(m
>= 0 && m
< board_size
&& n
>= 0 && n
<= board_size
);
hash_board(&prehash
, color
);
hash_board_and_move(&posthash
, color
, m
, n
);
if (collect_statistics
!= EMPTY
)
add_situation(&prehash
, &posthash
, collect_statistics
== color
,
color
== WHITE
? white_player
: black_player
);
store_pattern_if_winner(&prehash
, &posthash
, color
, m
, n
);
play_move(POS(m
, n
), color
);
fprintf(stderr
, "%8x ", prehash
.values
[l
]);
fprintf(stderr
, "%8x ", posthash
.values
[l
]);
fprintf(stderr
, "Node error\n");
/* Tests if the player has enough strength in the game to be interesting
enough_strength(char *strength
)
if (player_strength
>= 30)
length
= strlen(strength
);
/* check if dan or pro player */
for (i
= 0; i
< length
; i
++)
if (strength
[i
] == 'd' || strength
[i
] == 'D'
|| strength
[i
] == 'p' || strength
[i
] == 'P')
/* get the kyu strength as an integer */
for (i
= 0; i
< length
; i
++) {
if (player_strength
>= 30)
if (kyu
<= player_strength
)
/* not enough strength */
* used by both sort_games and collect_situations to
* check validity of a game record
* 0 means failure for any reason
* 1 means probably okay, without going through moves
check_game(SGFNode
*sgf
, char *sgfname
)
char *WR
, *BR
; /* white rank */
char thirty_kyu
[] = "30k";
if (!sgfGetIntProperty(sgf
, "SZ", &size
)) {
fprintf(stderr
, "Warning: no SZ in sgf file %s , assuming size %d\n",
if (size
!= board_size
) {
fprintf(stderr
, "Warning: wrong size of board %d in sgf file %s .\n",
if (handicap_value
== 0) {
if (sgfGetIntProperty(sgf
, "HA", &handicap
) && handicap
> 1) {
"No handicap games allowed, sgf file %s has handicap %d\n",
/* Only handicap games */
if (handicap_value
> 1) {
if (!sgfGetIntProperty(sgf
, "HA", &handicap
)) {
fprintf(stderr
, "Sgf file %s is not a handicap game\n", sgfname
);
/* only specific handicap games */
if (handicap_value
!= 10 && handicap
!= handicap_value
) {
"Sgf file %s has wrong number of handicap stones %d\n",
/* any reasonable handicap games */
if (handicap_value
== 10 && (handicap
< 2 || handicap
> 9)) {
"Sgf file %s has wrong/weird number of handicap stones %d\n",
/* Examine strength of players in the game and compare it
* with minimum player strength.
if (!sgfGetCharProperty(sgf
, "BR", &BR
)) {
fprintf(stderr
, "No black strength in sgf file %s assuming %s\n",
if (!enough_strength(BR
)) {
fprintf(stderr
, "Wrong black strength %s in sgf file %s\n", BR
, sgfname
);
if (!sgfGetCharProperty(sgf
, "WR", &WR
)) {
fprintf(stderr
, "No white strength in sgf file %s assuming %s\n",
if (!enough_strength(WR
)) {
fprintf(stderr
, "Wrong white strength %s in sgf file %s\n", WR
, sgfname
);
if (!sgfGetCharProperty(sgf
, "RE", &RE
)) {
fprintf(stderr
, "No result in game %s\n", sgfname
);
if (strncmp(RE
, "B+", 2) != 0 && strncmp(RE
, "W+", 2) != 0) {
fprintf(stderr
, "Couldn't parse winner in result %s from game %s\n",
* Sort out the games that can be used.
for (k
= 0; k
< number_of_games
; k
++) {
fprintf(stderr
, "Sorting number %d, %s\n", k
, sgf_names
[k
]);
sgf
= readsgffilefuseki(sgf_names
[k
], 0);
fprintf(stderr
, "Warning: Couldn't open sgf file %s number %d.\n",
unused_games
[k
] = 1; /* the game could not be used */
if (!check_game(sgf
, sgf_names
[k
]))
/* Free memory of SGF file */
/* Play through the initial moves of all games and collect hash values
* for the encountered situations.
int winner
; /* who won the game in question */
for (k
= 0; k
< number_of_games
; k
++) {
fprintf(stderr
, "Reading number %d, %s\n", k
, sgf_names
[k
]);
sgf
= readsgffilefuseki(sgf_names
[k
], moves_per_game
);
fprintf(stderr
, "Warning: Couldn't open sgf file %s.\n", sgf_names
[k
]);
unused_games
[k
] = 1; /* the game could not be used */
if (!check_game(sgf
, sgf_names
[k
])) {
if (!sgfGetCharProperty(sgf
, "RE", &RE
)) {
if (strncmp(RE
, "B+", 2) == 0)
else if (strncmp(RE
, "W+", 2) == 0)
if (!examine_game(sgf
, winner
)) {
fprintf(stderr
, "Warning: Problem with sgf file %s\n", sgf_names
[k
]);
unused_games
[k
] = 1; /* the game could not be used */
/* Free memory of SGF file */
/* Find the most common positions and moves, for which we want to
/* Sort all the collected situations. */
gg_sort(situation_table
, number_of_situations
, sizeof(*situation_table
),
for (i
= 0; i
< number_of_situations
; i
++) {
fprintf(stderr
, "%4d ", i
);
fprintf(stderr
, "%8x ", situation_table
[i
].pre
.values
[k
]);
fprintf(stderr
, "%8x ", situation_table
[i
].post
.values
[k
]);
/* Set up frequency table. */
frequency_table
= calloc(number_of_situations
, sizeof(*frequency_table
));
fprintf(stderr
, "Fatal error, failed to allocate frequency table.\n");
number_of_distinct_positions
= 0;
/* Make frequency analysis of the positions before the moves. */
for (k
= 0; k
< number_of_situations
; k
++) {
if (k
== 0 || compare_positions(&situation_table
[k
],
&situation_table
[k
-1]) != 0) {
frequency_table
[number_of_distinct_positions
].index
= k
;
frequency_table
[number_of_distinct_positions
].n
= 0;
frequency_table
[number_of_distinct_positions
].n_win
= 0;
frequency_table
[number_of_distinct_positions
].n_player
= 0;
number_of_distinct_positions
++;
frequency_table
[number_of_distinct_positions
-1].n
++;
frequency_table
[number_of_distinct_positions
-1].n_win
+= situation_table
[k
].outcome
;
if (frequency_table
[number_of_distinct_positions
-1].n
== 1
|| situation_table
[k
].player
!= situation_table
[k
-1].player
)
frequency_table
[number_of_distinct_positions
-1].n_player
++;
/* Sort the frequency table, in falling order. */
gg_sort(frequency_table
, number_of_distinct_positions
,
sizeof(*frequency_table
), compare_frequencies
);
for (l
= 0; l
< number_of_distinct_positions
; l
++) {
fprintf(stderr
, "%4d %5d\n", frequency_table
[l
].n
,
frequency_table
[l
].index
);
/* Set up winners array. */
winning_moves
= calloc(MAX_PATTERNS_TO_EXTRACT
, sizeof(*winning_moves
));
fprintf(stderr
, "Fatal error, failed to allocate winning moves table.\n");
number_of_winning_moves
= 0;
/* Starting with the most common position, find the most common
* moves for each position, until the number of patterns to be
for (k
= 0; k
< number_of_distinct_positions
; k
++) {
int index
= frequency_table
[k
].index
;
/* Build a new frequency table for the different moves in this position. */
struct frequency move_frequencies
[MAX_BOARD
* MAX_BOARD
];
/* A position must occur a minimum before we analyze and record
if (frequency_table
[k
].n
< min_position_freq
)
if (i
== number_of_situations
&& compare_positions(&situation_table
[i
],
&situation_table
[i
-1]) != 0))
if (i
== index
|| compare_situations(&situation_table
[i
],
&situation_table
[i
-1]) != 0) {
move_frequencies
[number_of_moves
].index
= i
;
move_frequencies
[number_of_moves
].n
= 0;
move_frequencies
[number_of_moves
].n_win
= 0;
move_frequencies
[number_of_moves
].n_player
= 0;
move_frequencies
[number_of_moves
-1].n
++;
move_frequencies
[number_of_moves
-1].n_win
+= situation_table
[i
].outcome
;
if (move_frequencies
[number_of_moves
-1].n
== 1
|| situation_table
[i
].player
!= situation_table
[i
-1].player
)
move_frequencies
[number_of_moves
-1].n_player
++;
/* Sort the moves, in falling order. */
gg_sort(move_frequencies
, number_of_moves
,
sizeof(*move_frequencies
), compare_frequencies2
);
for (i
= 0; i
< number_of_moves
; i
++) {
fprintf(stderr
, "%4d %3d %4d\n", index
, move_frequencies
[i
].n
,
move_frequencies
[i
].index
);
/* Add the moves to the list of winners. */
for (i
= 0; i
< number_of_moves
; i
++) {
/* This is where the cut-off of moves is decided
* based on popularity from command line arguments.
if (0.01 * min_move_percent
*move_frequencies
[0].n_player
> move_frequencies
[i
].n_player
|| move_frequencies
[i
].n_player
< min_move_freq
) {
winning_moves
[number_of_winning_moves
].index
= -1;
winning_moves
[number_of_winning_moves
].pre
=
situation_table
[frequency_table
[k
].index
].pre
.values
[0];
winning_moves
[number_of_winning_moves
].position_frequency
=
winning_moves
[number_of_winning_moves
].n_player
= 0;
winning_moves
[number_of_winning_moves
].move_frequency
= 0;
winning_moves
[number_of_winning_moves
].position_success
=
frequency_table
[k
].n_win
;
winning_moves
[number_of_winning_moves
].move_success
= 0;
while (i
< number_of_moves
) {
gg_assert(0.01 * min_move_percent
*move_frequencies
[0].n_player
> move_frequencies
[i
].n_player
|| move_frequencies
[i
].n_player
< min_move_freq
);
gg_assert(situation_table
[move_frequencies
[i
].index
].pre
.values
[0]
== winning_moves
[number_of_winning_moves
].pre
);
winning_moves
[number_of_winning_moves
].n_player
+=
move_frequencies
[i
].n_player
;
winning_moves
[number_of_winning_moves
].move_frequency
+=
winning_moves
[number_of_winning_moves
].move_success
+=
move_frequencies
[i
].n_win
;
number_of_winning_moves
++;
winning_moves
[number_of_winning_moves
].index
= move_frequencies
[i
].index
;
winning_moves
[number_of_winning_moves
].pre
=
situation_table
[frequency_table
[k
].index
].pre
.values
[0];
winning_moves
[number_of_winning_moves
].position_frequency
=
winning_moves
[number_of_winning_moves
].move_frequency
=
winning_moves
[number_of_winning_moves
].n_player
=
move_frequencies
[i
].n_player
;
winning_moves
[number_of_winning_moves
].position_success
=
frequency_table
[k
].n_win
;
winning_moves
[number_of_winning_moves
].move_success
=
move_frequencies
[i
].n_win
;
number_of_winning_moves
++;
if (number_of_winning_moves
== MAX_PATTERNS_TO_EXTRACT
)
if (number_of_winning_moves
== MAX_PATTERNS_TO_EXTRACT
)
for (i
= 0; i
< number_of_winning_moves
; i
++) {
fprintf(stderr
, "%4d %3d %3d\n",
winning_moves
[i
].position_frequency
,
winning_moves
[i
].move_frequency
);
/* Scan through the games a second time to pick up the patterns
* corresponding to the winning moves.
for (k
= 0; k
< number_of_games
; k
++) {
fprintf(stderr
, "Generating number %d, %s\n", k
, sgf_names
[k
]);
/* Check if this game is a valid game. */
fprintf(stderr
, "Not used\n");
sgf
= readsgffilefuseki(sgf_names
[k
], moves_per_game
);
fprintf(stderr
, "Warning: Couldn't open sgf file %s.\n", sgf_names
[k
]);
/* Free memory of SGF file. */
/* Print the winning patterns in patterns.db format on stdout. */
unsigned int pre
= situation_table
[winning_moves
[0].index
].pre
.values
[0];
gg_assert(winning_moves
[0].index
!= -1);
for (k
= 0; k
< number_of_winning_moves
; k
++) {
/* Do not print erroneous patterns. */
if (winning_moves
[k
].pattern
[0][0] != '\0'
|| winning_moves
[k
].index
== -1) {
double grand_sum
= winning_moves
[k
].position_frequency
;
double grand_wins
= winning_moves
[k
].position_success
;
double grand_losses
= grand_sum
- grand_wins
;
double row_sum
= winning_moves
[k
].move_frequency
;
double wins
= winning_moves
[k
].move_success
;
double losses
= row_sum
- wins
;
double expect_wins
= row_sum
*grand_wins
/grand_sum
;
double expect_losses
= row_sum
- expect_wins
;
/* We're _not_ using a Yates corrected chisquare.
* Two reasons: 1. It's never correct for > 2x2 table
* 2. Our marginals are sampled, not fixed, so
* Yates and usual Fisher exact are wrong distribution.
* Straight chi squared is best.
/* This was Yates line. It's wrong. */
dchisq
+= pow(gg_abs(wins
- expect_wins
) - 0.5, 2) / expect_wins
;
dchisq
+= pow(wins
- expect_wins
, 2) / expect_wins
;
dchisq
+= pow(losses
- expect_losses
, 2) / expect_losses
;
gg_assert(winning_moves
[k
].index
== -1
|| (situation_table
[winning_moves
[k
].index
].pre
.values
[0]
== winning_moves
[k
].pre
));
/* Did we just finish a set? If so, print totals and reset. */
if (winning_moves
[k
].pre
!= pre
) {
/* p-value is 1 - incomplete gamma function(d.o.f/2, chisq/2)
* variable df is number of moves, actual d.o.f is df-1
* k is referring to the pattern _after_ the set we just completed.
printf("\n### Summary of pattern pre 0x%08x\n### N Chi_squared df: %d %g %d ",
pre
, winning_moves
[k
-1].position_frequency
, chisq
, df
- 1);
/* and array is indexed at zero for d.o.f = 1... */
else if (df
- 1 < (int) (sizeof(chisquarecrit05
) / sizeof(double))
&& chisq
> chisquarecrit05
[df
-2]) {
/* The overall result is significant at 5%. In this case, at
* least some authorities will allow us to examine several
* individual contrasts w/o futher correction. We compare
* every pair of moves, which is perhaps too many, but the
* purpose is to give the human expert (who would in any
* case be required to examine the output) some sense of
* where the differences are. Something like a Bonferroni
* correction could result in a significant test overall,
* but no significant contrasts, which is obviously
* nonsense. The significance of the overall result must
if (chisq
> chisquarecrit001
[df
-2])
printf("!!! p < 0.001\n");
else if (chisq
> chisquarecrit01
[df
-2])
printf("!!! p < 0.01\n");
printf("!!! p < 0.05\n");
for (m
= first_in_set
; m
< k
; m
++) {
for (n
= m
+ 1; n
< k
; n
++) {
double grand_sum
= (winning_moves
[m
].move_frequency
+ winning_moves
[n
].move_frequency
);
double grand_wins
= (winning_moves
[m
].move_success
+ winning_moves
[n
].move_success
);
double grand_losses
= grand_sum
- grand_wins
;
double row_sum_m
= winning_moves
[m
].move_frequency
;
double row_sum_n
= winning_moves
[n
].move_frequency
;
double wins_m
= winning_moves
[m
].move_success
;
double losses_m
= row_sum_m
- wins_m
;
double wins_n
= winning_moves
[n
].move_success
;
double losses_n
= row_sum_n
- wins_n
;
double expect_wins_m
= row_sum_m
* grand_wins
/grand_sum
;
double expect_losses_m
= row_sum_m
- expect_wins_m
;
double expect_wins_n
= row_sum_n
* grand_wins
/grand_sum
;
double expect_losses_n
= row_sum_n
- expect_wins_n
;
dchisq_m
+= pow(wins_m
- expect_wins_m
, 2) / expect_wins_m
;
if (expect_losses_m
> 0.0)
dchisq_m
+= pow(losses_m
- expect_losses_m
, 2) / expect_losses_m
;
dchisq_n
+= pow(wins_n
- expect_wins_n
, 2) / expect_wins_n
;
if (expect_losses_n
> 0.0)
dchisq_n
+= pow(losses_n
- expect_losses_n
, 2) / expect_losses_n
;
/* We demand at least N=6. Nonsense with smaller N. */
if (dchisq_m
+ dchisq_n
> chisquarecrit05
[0] && grand_sum
> 5) {
printf("#### 0x%08x %c 0x%08x (p < 0.05) chisq = %g N = %g\n",
situation_table
[winning_moves
[m
].index
].post
.values
[0],
(1.0 * wins_m
/ row_sum_m
> 1.0 * wins_n
/ row_sum_n
) ? '>' : '<',
situation_table
[winning_moves
[n
].index
].post
.values
[0],
dchisq_m
+ dchisq_n
, grand_sum
);
else if (df
-1 < (int) (sizeof(chisquarecrit10
) / sizeof(double))
&& chisq
> chisquarecrit10
[df
- 2])
printf("??? p < 0.10\n\n");
else if (!(df
- 1 < (int) (sizeof(chisquarecrit05
) / sizeof(double))))
printf("df out of range...\n\n");
pre
= winning_moves
[k
].pre
;
pre
= situation_table
[winning_moves
[k
].index
].pre
.values
[0];
if (winning_moves
[k
].index
== -1) {
printf("# Unpopular moves\n");
printf("# pre: 0x%08x\n", winning_moves
[k
].pre
);
printf("# post: could be various\n");
printf("# frequency: %d/%d\n",
winning_moves
[k
].move_frequency
,
winning_moves
[k
].position_frequency
);
printf("# unique players: %d\n", winning_moves
[k
].n_player
);
printf("# wins: %d/%d\n\n",
winning_moves
[k
].move_success
,
winning_moves
[k
].position_success
);
printf("# success: %.1f%% vs %.1f%% for this position overall, dchisq = %g\n\n",
100.0 * winning_moves
[k
].move_success
/ winning_moves
[k
].move_frequency
,
100.0 * winning_moves
[k
].position_success
/ winning_moves
[k
].position_frequency
,
printf("Pattern F-H%d-%d\n", handicap_value
, l
);
printf("# pre : 0x%08x\n",
situation_table
[winning_moves
[k
].index
].pre
.values
[0]);
printf("# post: 0x%08x\n",
situation_table
[winning_moves
[k
].index
].post
.values
[0]);
printf("# frequency: %d/%d\n", winning_moves
[k
].move_frequency
,
winning_moves
[k
].position_frequency
);
printf("# unique players: %d\n", winning_moves
[k
].n_player
);
printf("# wins: %d/%d\n\n", winning_moves
[k
].move_success
,
winning_moves
[k
].position_success
);
printf("# success: %.1f%% vs %.1f%% for this position overall, dchisq = %g\n\n",
100.0 * winning_moves
[k
].move_success
/ winning_moves
[k
].move_frequency
,
100.0 * winning_moves
[k
].position_success
/ winning_moves
[k
].position_frequency
,
for (n
= 0; n
< board_size
; n
++)
for (m
= 0; m
< board_size
; m
++) {
for (n
= 0; n
< board_size
; n
++) {
if (winning_moves
[k
].pattern
[m
][n
] == '\0') {
fprintf(stderr
, "Something wrong in print pattern\n");
printf("%c", winning_moves
[k
].pattern
[m
][n
]);
for (n
= 0; n
< board_size
; n
++)
printf("\n\n:8,-,value(%d)\n\n\n", winning_moves
[k
].n_player
);
"Skipping pattern pre 0x%08x post 0x%08x, stats may be wrong...\n",
situation_table
[winning_moves
[k
].index
].pre
.values
[0],
situation_table
[winning_moves
[k
].index
].post
.values
[0]);
main(int argc
, char *argv
[])
int number_of_unused_games
= 0;
/* Check number of arguments. */
board_size
= atoi(argv
[2]);
if (board_size
% 2 == 0) {
fprintf(stderr
, "Fatal error, only odd boardsizes supported: %d.\n",
if (board_size
< 9 || board_size
> 19)
fprintf(stderr
, "Warning: strange boardsize: %d.\n", board_size
);
moves_per_game
= atoi(argv
[3]);
if (moves_per_game
< 1 || moves_per_game
> 20)
fprintf(stderr
, "Warning: strange number of moves per game: %d.\n",
handicap_value
= atoi(argv
[4]);
if (handicap_value
< 0 || handicap_value
> 10)
fprintf(stderr
, "Warning: unusual handicap value: %d.\n",
player_strength
= atoi(argv
[5]);
if (player_strength
< 0 || player_strength
> 30)
fprintf(stderr
, "Warning: wrong lowest strength: %d.\n",
half_board_patterns
= atoi(argv
[6]);
if (half_board_patterns
!= 0 && half_board_patterns
!= 1) {
"Warning: incorrect half_board_flag (0 or 1). Setting the value to 0.\n");
min_position_freq
= atoi(argv
[7]);
if (min_position_freq
< 1) {
fprintf(stderr
, "Warning: setting min_position_freq to 1\n");
min_move_percent
= atof(argv
[8]);
if (min_move_percent
< 0. || min_move_percent
> 100.) {
fprintf(stderr
, "Warning: strange min_move_percent %g, setting to 1%%\n",
min_move_freq
= atoi(argv
[9]);
fprintf(stderr
, "Warning: strange min_move_freq %d\n", min_move_freq
);
/* Count the number of sgf files. */
number_of_games
= read_sgf_filenames(argv
[1], NULL
);
/* Allocate space for the list of unused files. */
unused_games
= calloc(number_of_games
, sizeof(*unused_games
));
if (unused_games
== NULL
) {
fprintf(stderr
, "Fatal error, failed to allocate memory.\n");
/* Allocate space for the list of sgf file names. */
sgf_names
= calloc(number_of_games
, sizeof(*sgf_names
));
fprintf(stderr
, "Fatal error, failed to allocate memory.\n");
/* Read the list of sgf files and store in memory. */
read_sgf_filenames(argv
[1], sgf_names
);
/* Save memory by sorting out the games that can be used first */
fprintf(stderr
, "Starting game sort\n");
fprintf(stderr
, "Starting game writes\n");
write_sgf_filenames(argv
[10], sgf_names
);
/* Build tables of random numbers for Zobrist hashing. */
/* Play through the initial moves of all games and collect hash values
* for the encountered situations.
fprintf(stderr
, "collect OK.\n");
/* Find the most common positions and moves, for which we want to
fprintf(stderr
, "analyze OK.\n");
/* Generate patterns from the chosen positions and moves.
fprintf(stderr
, "generate OK.\n");
printf("attribute_map value_only\n\n\n");
for (i
= 0; i
< argc
; i
++)
/* Write the patterns to stdout in patterns.db format.
/* Tell the user everything worked out fine */
fprintf(stderr
, "The pattern database was produced with no errors.\n");
for (i
= 0; i
< number_of_games
; i
++)
number_of_unused_games
++;
fprintf(stderr
, "Out of %d games, %d were not used.\n",
number_of_games
, number_of_unused_games
);