[ << ] [ >> ]           [Top] [Contents] [Index] [ ? ]

13. Influence Function

13.1 Conceptual Outline of Influence  
13.2 Territory, Moyo and Area  
13.3 Where influence gets used in the engine  
13.4 Influence and Territory  
13.5 Details of the Territory Valuation  
13.6 The Core of the Influence Function  
13.7 The Influence Algorithm  The algorithm of accumlate_influence()
13.8 Permeability  
13.9 Escape  
13.10 Break Ins  
13.11 Surrounded Dragons  
13.12 Patterns used by the Influence module  
13.13 Colored display and debugging of influence  
13.14 Influence Tuning with view.pike  Influence tuning with view.pike


13.1 Conceptual Outline of Influence

We define call stones lively if they cannot be tactically attacked, or if they have a tactical defense and belong to the player whose turn it is. Similarly, stones that cannot be strategically attacked (in the sense of the life-and-death analysis), or that have a strategical defense and belong to the player to move, are called alive. If we want to use the influence function before deciding the strategical status, all lively stones count as alive.

Every alive stone on the board works as an influence source, with influence of its color radiating outwards in all directions. The strength of the influence declines exponentially with the distance from the source.

Influence can only flow unhindered if the board is empty, however. All lively stones (regardless of color) act as influence barriers, as do connections between enemy stones that can't be broken through. For example the one space jump counts as a barrier unless either of the stones can be captured. Notice that it doesn't matter much if the connection between the two stones can be broken, since in that case there would come influence from both directions anyway.

From the influence of both colors we compute a territorial value between -1.0 and +1.0 for each intersection, which can be seen as the likely hood of it becoming territory for either color.

In order to avoid finding bogus territory, we add extra influence sources at places where an invasion can be launched, e.g. at 3-3 under a handicap stone, in the middle of wide edge extensions and in the center of large open spaces anywhere. Similarly we add extra influence sources where intrusions can be made into what otherwise looks as solid territory, e.g. monkey jumps. These intrusions depend on whose turn we assume it to be.

All these extra influence sources, as well as connections, are controlled by a pattern database, which consists of the two files patterns/influence.db and patterns/barriers.db. The details are explained in 13.12 Patterns used by the Influence module.


13.2 Territory, Moyo and Area

Using the influence code, empty regions of the board are partitioned in three ways. A vertex may be described as White or Black's territory, moyo or area. The functions whose_territory(), whose_moyo() and whose_area() will return a color, or EMPTY if it belongs to one player or the other in one of these classifications.

Generally territory is moyo and moyo is area. To get a feeling for these concepts, load an sgf file in a middle game position with the option `-m 0x0180' and examine the resulting diagrams (see section 13.13 Colored display and debugging of influence).


13.3 Where influence gets used in the engine

The information obtained from the influence computation is used in a variety of places in the engine, and the influence module is called several times in the process of the move generation. The details of the influence computation vary according to the needs of the calling function.

After GNU Go has decided about the tactical stability of strings, the influence module gets called the first time. Here all lively stones act as an influence source of default strength 100. The result is stored in the variables initial_influence and initial_opposite_influence, and it is used as an important information for guessing the strength of dragons. For example, a dragon that is part of a moyo of size 25 is immediately considered alive. For dragons with a smaller moyo size, a life-and-death analysis will be done by the owl code (see 12. Pattern Based Reading). A dragon with a moyo size of only 5 will be considered weak, even if the owl code has decided that it cannot be killed.

As a tool for both the owl code and the strength estimate of dragons, an "escape" influence gets computed for each dragon (see section 13.9 Escape).

Once all dragons have been evaluated, the influence module is called again and the variables initial_influence and initial_opposite_influence get overwritten. Of course, the dragon status', which are available now, are taken into account. Stones belonging to a dead dragon will not serve as an influence source, and the strengths of other stones get adjusted according to the strength of their respective dragon.

The result of this run is the most important tool for move evaluation. All helper functions of patterns as explained in 9. The Pattern Code that refer to influence results (e. g. olib(*) etc.) actually use these results. Further, initial_influence serves as the reference for computing the territorial value of a move. That is, from the influence strengths stored in initial_influence, a territory value is assigned to each intersection. This value is supposed to estimate the likelyhood that this intersection will become white or black territory.

Then, for each move that gets considered in the function value_moves, the influence module is called again via the function compute_move_influence to assess the likely territorial balance after this move, and the result is compared with the state before that move.

An additional influence computation is done in order to compute the followup value of a move. Some explainations are in 13.5 Details of the Territory Valuation.

Some of the public functions from `influence.c' which are used throughout the engine are listed in 18.4 Utilities from `engine/influence.c'.


13.4 Influence and Territory

In this section we consider how the influence function is used to estimate territory in the function estimate_territorial_value().

A move like `*' by `O' below is worth one point:

 
OXXX.
OX.XX
O*a.X
OX.XX
OXXX.

This is evaluated by the influence function in the following way: We first assign territory under the assumption that X moves first in all local positions in the original position; then we reassing territory, again under the assumption that `X' moves first in all local positions, but after we let `O' make the move at `*'. These two territory assignments are compared and the difference gives the territorial value of the move.

Technically, the assumption that `X' plays first everywhere is implemented via an asymmetric pattern database in barriers.db. What exactly is a safe connection that stops hostile influence from passing through is different for `O' and `X'; of course such a connection has to be tighter for stones with color `O'. Also, additional intrusion influence sources are added for `X' in places where `X' stones have natural followup moves.

In this specific example above, the asymmetry (before any move has been made) would turn out as follows: If `X' is in turn to move, the white influence would get stopped by a barrier at `*', leaving 4 points of territory for `X'. However, if `O' was next to move, then a followup move for the white stones at the left would be assumed in the form of an extra ("intrusion") influence source at `*'. This would get stopped at `a', leaving three points of territory.

Returning to the valuation of a move by `O' at `*', we get a value of 1 for the move at `*'. However, of course this move is sente once it is worth playing, and should therefore (in miai counting) be awarded an effective value of 2. Hence we need to recognize the followup value of a move. GNU Go 3.0 took care of this by using patterns in patterns.db that enforced an explicit followup value. Versions from 3.2 on instead compute a seperate followup influence to each move considered. In the above example, an intrusion source will be added at `a' as a followup move to `*'. This destroys all of Black's territory and hence gives a followup value of 3.

The pattern based followup value are still needed at some places, however.

To give another example, consider this position where we want to estimate the value of an `O' move at `*':

 
OOOXXX
..OX..
..OX..
...*..
------

Before the move we assume `X' moves first in the local position (and that `O' has to connect), which gives territory like this (lower case letter identify territory for each player):

 
OOOXXX
ooOXxx
o.OXxx
o...xx
------

Then we let `O' make the move at `*' and assume `X' moves first again next. The territory then becomes (`X' is also assumed to have to connect):

 
OOOXXX
ooOXxx
ooOX.x
oo.O.x
------

We see that this makes a difference in territory of 4, which is what influence_delta_territory() should report. Then again, we have followup value, and here also a reverse followup value. The reverse followup value, which in this case will be so high that the move is treated as reverse sente, is added by an explicit pattern. Other sources for followup or reverse followup values are threats to capture a rescue a string of stones. See the code and comments in the function value_move_reaons for how followup and reverse followup values are used to adjust the effective move value.

To give an example of territorial value where something is captured, consider the `O' move at `*' here,

 
XXXXXXXO
X.OOOOXO
X.O..O*O
--------

As before we first let the influence function determine territory assuming X moves first, i.e. with a captured group:

 
XXXXXXXO
XxyyyyXO
Xxyxxy.O
--------

Here `y' indicates `X' territory + captured stone, i.e. these count for two points. After the `O' move at `*' we instead get

 
XXXXXXXO
X.OOOOXO
X.OooOOO
--------

and we see that `X' has 16 territory fewer and `O' has two territory more, for a total difference of 18 points.

That the influence function counts the value of captured stones was introduced in GNU Go 3.2. Previously this was instead done using the effective_size heuristic. The effective size is the number of stones plus the surrounding empty spaces which are closer to this string or dragon than to any other stones. Here the `O' string would thus have effective size 6 (number of stones) + 2 (interior eye) + 2*0.5 (the two empty vertices to the left of the string, split half each with the surrounding X string) + 1*0.33 (the connection point, split between three strings) = 9.33. As noted this value was doubled, giving 18.67 which is reasonably close to the correct value of 18. The effective size heuristic is still used in certain parts of the move valuation where we can't easily get a more accurate value from the influence function (e. g. attacks depending on a ko, attack threats).

Note that this section only describes the territorial valuation of a move. Apart from that, GNU Go uses various heuristics in assigning a strategical value (weakening and strengthening of other stones on the board) to a move. Also, the influence function isn't quite as well tuned as the examples above may seem to claim. But it should give a fairly good idea of how the design is intended.

Another matter is that so far we have only considered the change in secure territory. GNU Go 3.2 and later versions use a revised heuristic, which is explained in the next section, to assign probable territory to each player.


13.5 Details of the Territory Valuation

This section explains how GNU Go assigns a territorial value to an intersection once the white and black influence have been computed. The intention is that an intersection that has a chance of xx% of becoming white territory is counted as 0.xx points of territory for white, and similar for black.

The algorithm in the function new_value_territory goes roughly as follows:

If wi is the white influence at a point, and bi the black influence, then value = ( (wi-bi)/ (wi+bi) )^3 (positive values indicates likley white territory, negative stand for black territory) turns out to be very simple first guess that is still far off, but reasonable enough to be useful.

This value is then suspect a number of corrections. Assume that this first guess resulted in a positive value.

If both bi and wi are small, it gets reduced. What exactly is "small" depends on whether the intersection is close to a corner or an edge of the board, since it is easier to claim territory in the corner than in the center.

Then the value at each intersection is degraded to the minimum value of its neighbors. This can be seen as a second implementation of the proverb saying that there is no territory in the center of the board. This step substantially reduces the size of spheres of territory that are open at several sides.

Finally, there are a number of patterns that explicitly forbid GNU Go to count territory at some intersections. This is used e. g. for false eyes that will eventually have to be filled in. Also, points for prisoners are added.

To fine tune this scheme, some revisions have been made to the influence computations that are relevant for territorial evaluation. This includes a reduced default attenuation and some revised pattern handling.


13.6 The Core of the Influence Function

The basic influence radiation process can efficiently be implemented as a breadth first search for adjacent and more distant points, using a queue structure.

Influence barriers can be found by pattern matching, assisted by reading through constraints and/or helpers. Wall structures, invasion points and intrusion points can be found by pattern matching as well.

When influence is computed, the basic idea is that there are a number of influence sources on the board, whose contributions are summed to produce the influence values. For the time being we can assume that the living stones on the board are the influence sources, although this is not the whole story.

The function compute_influence() contains a loop over the board, and for each influence source on the board, the function accumulate_influence() is called. This is the core of the influence function. Before we get into the details, this is how the influence field from a single isolated influence source of strength 100 turns out (with an attenuation of 3.0):

 
  0  0  0  0  0  0  0  0  0  0  0
  0  0  0  0  1  1  1  0  0  0  0
  0  0  0  1  2  3  2  1  0  0  0
  0  0  1  3  5 11  5  3  1  0  0
  0  1  2  5 16 33 16  5  2  1  0
  0  1  3 11 33  X 33 11  3  1  0
  0  1  2  5 16 33 16  5  2  1  0
  0  0  1  3  5 11  5  3  1  0  0
  0  0  0  1  2  3  2  1  0  0  0
  0  0  0  0  1  1  1  0  0  0  0
  0  0  0  0  0  0  0  0  0  0  0

These values are in reality floating point numbers but have been rounded down to the nearest integer for presentation. This means that the influence field does not stop when the numbers become zeroes.

Internally accumulate_influence() starts at the influence source and spreads influence outwards by means of a breadth first propagation, implemented in the form of a queue. The order of propagation and the condition that influence only is spread outwards guarantee that no intersection is visited more than once and that the process terminates. In the example above, the intersections are visited in the following order:

 
  +  +  +  +  +  +  +  +  +  +  +
  + 78 68 66 64 63 65 67 69 79  +
  + 62 46 38 36 35 37 39 47 75  +
  + 60 34 22 16 15 17 23 43 73  +
  + 58 32 14  6  3  7 19 41 71  +
  + 56 30 12  2  0  4 18 40 70  +
  + 57 31 13  5  1  8 20 42 72  +
  + 59 33 21 10  9 11 24 44 74  +
  + 61 45 28 26 25 27 29 48 76  +
  + 77 54 52 50 49 51 53 55 80  +
  +  +  +  +  +  +  +  +  +  +  +

The visitation of intersections continues in the same way on the intersections marked '`+' and further outwards. In a real position there will be stones and tight connections stopping the influence from spreading to certain intersections. This will disrupt the diagram above, but the main property of the propagation still remains, i.e. no intersection is visited more than once and after being visited no more influence will be propagated to the intersection.


13.7 The Influence Algorithm

Let (m, n) be the coordinates of the influence source and (i, j) the coordinates of a an intersection being visited during propagation, using the same notation as in the accumulate_influence() function. Influence is now propagated to its eight closest neighbors, including the diagonal ones, according to the follow scheme:

For each of the eight directions (di, dj), do:

  1. Compute the scalar product di*(i-m) + dj*(j-n) between the vectors (di,dj) and (i,j) - (m,n)
  2. If this is negative or zero, the direction is not outwards and we continue with the next direction. The exception is when we are visiting the influence source, i.e. the first intersection, when we spread influence in all directions anyway.
  3. If (i+di, j+dj) is outside the board or occupied we also continue with the next direction.
  4. Let S be the strength of the influence at (i, j). The influence propagated to (i+di, j+dj) from this intersection is given by P*(1/A)*D*S, where the three different kinds of damping are:

Influence is typically contributed from up to three neighbors "between" this intersection and the influence source. These values are simply added together. As pointed out before, all contributions will automatically have been made before the intersection itself is visited.

When the total influence for the whole board is computed by compute_influence(), accumulate_influence() is called once for each influence source. These invocations are totally independent and the influence contributions from the different sources are added together.


13.8 Permeability

The permeability at the different points is initially one at all empty intersections and zero at occupied intersections. To get a useful influence function we need to modify this, however. Consider the following position:

 
|......
|OOOO..
|...O..
|...a.X   ('a' empty intersection)
|...O..
|...OOO
|.....O
+------

The corner is of course secure territory for `O' and clearly the `X' stone has negligible effect inside this position. To stop `X' influence from leaking into the corner we use pattern matching (pattern Barrier1/Barrier2 in `barriers.db') to modify the permeability for `X' at this intersection to zero. `O' can still spread influence through this connection.

Another case that needs to be mentioned is how the permeability damping is computed for diagonal influence radiation. For horizontal and vertical radiation we just use the permeability (for the relevant color) at the intersection we are radiating from. In the diagonal case we additionally multiply with the maximum permeability at the two intersections we are trying to squeeze between. The reason for this can be found in the diagram below:

 
|...X    |...X    
|OO..    |Oda.
|..O.    |.bc.
|..O.    |..O.
+----    +----

We don't want `X' influence to be spread from `a' to `b', and since the permeability at both c and d is zero, the rule above stops this.


13.9 Escape

One application of the influence code is in computing the dragon.escape_route field. This is computed by the function compute_escape() as follows. First, every intersection is assigned an escape value, ranging between 0 and 4, depending on the influence value of the opposite color.

The escape_route field is modified by the code in `surround.c' (see section 13.11 Surrounded Dragons). It is divided by two for weakly surrounded dragons, and set to zero for surrounded ones.

In addition to assiging an escape value to empty vertices, we also assign an escape value to friendly dragons. This value can range from 0 to 6 depending on the status of the dragon, with live dragons having value 6.

Then we sum the values of the resulting influence escape values over the intersections (including friendly dragons) at distance 4, that is, over those intersections which can be joined to the dragon by a path of length 4 (and no shorter path) not passing adjacent to any unfriendly dragon. In the following example, we sum the influence escape value over the four vertices labelled '4'.

 
   
   . . . . . . . . .    . . . . . . . . .
   . . . . . X . . O    . . . . . X . . O
   . . X . . . . . O    . . X . 2 . 4 . O
   X . . . . . . . .    X . . 1 1 2 3 4 .
   X O . O . . . . O    X O 1 O 1 2 3 4 O
   X O . O . . . . .    X O 1 O 1 . 4 . .
   X O . . . X . O O    X O 1 . . X . . O
   . . . X . . . . .    . 1 . X . . . . .
   X . . . . X . . .    X . . . . X . . .
   . . . . . . . . .    . . . . . . . . .

Since the dragon is trying to reach safety, the reader might wonder why compute_influence() is called with the opposite color of the dragon contemplating escape. To explain this point, we first remind the reader why there is a color parameter to compute_influence(). Consider the following example position:
 
     ...XX...
     OOO..OOO
     O......O
     O......O
     --------

Whether the bottom will become O territory depends on who is in turn to play. This is implemented with the help of patterns in barriers.db, so that X influence is allowed to leak into the bottom if X is in turn to move but not if O is. There are also "invade" patterns which add influence sources in sufficiently open parts of the board which are handled differently depending on who is in turn to move.

In order to decide the territorial value of an O move in the third line gap above, influence is first computed in the original position with the opponent (i.e. X) in turn to move. Then the O stone is played to give:

 
     ...XX...
     OOO.OOOO
     O......O
     O......O
     --------

Now influence is computed once more, also this time with X in turn to move. The difference in territory (as computed from the influence values) gives the territorial value of the move.

Exactly how influence is computed for use in the escape route estimation is all ad hoc. But it makes sense to assume the opponent color in turn to move so that the escape possibilities aren't overestimated. After we have made a move in the escape direction it is after all the opponent's turn.

The current escape route mechanism seems good enough to be useful but is not completely reliable. Mostly it seems to err on the side of being too optimistic.


13.10 Break Ins

The code in `breakin.c' break-ins into territories that require deeper tactical reading and are thus impossible to detect for the influence module. It gets run after the influence module and revises its territory valuations.

The break-in code makes use of two public functions in `readconnect.c',

These functions are public front ends to their counterparts recursive_break_in and recursive_block_off, which call each other recursively.

The procedure is as follows: We look at all big (>= 10) territory regions as detected by the influence code. Using the computation of connection distances from readconnect.c, we compute all nearby vertices of this territory. We look for the closest safe stones belonging to the opponent.

For each such string str we call

If the break in is successful resp. the blocking unsuccessful, we shrink the territory, and see whether the opponent can still break in. We repeat this until the territory is shrunk so much that the opponent can no longer reach it.

To see the break in code in action run GNU Go on the file `regression/games/break_in.sgf' with the option -d0x102000. Among the traces you will find:

 
  Trying to break in from D7 to:
E9 (1)  F9 (1)  G9 (1)  E8 (1)  F8 (1)  G8 (1)  
H8 (1)  G7 (1)  H7 (1)  J7 (1)  H6 (1)  J6 (1)
H5 (1)  J5 (1)  H4 (1)  J4 (1)  H3 (1)  J3 (1)
H2 (1)  J2 (1)    
block_off D7, result 0 PASS (355, 41952 nodes, 0.73 seconds)
E9 (1)  F9 (1)  G9 (1)  E8 (1)  F8 (1)  G8 (1)
H8 (1)  G7 (1)  H7 (1)  J7 (1)  H6 (1)  J6 (1)
H5 (1)  J5 (1)  H4 (1)  J4 (1)  H3 (1)  J3 (1)
H2 (1)  J2 (1)    
B:F4 
  Erasing territory at E8 -b.
  Erasing territory at G3 -b.
  Now trying to break to smaller goal:
F9 (1)  G9 (1)  F8 (1)  G8 (1)  H8 (1)  G7 (1)
H7 (1)  J7 (1)  H6 (1)  J6 (1)  H5 (1)  J5 (1)
H4 (1)  J4 (1)  H3 (1)  J3 (1)  H2 (1)  J2 (1)    

This means that the function break_in is called with the goal marked 'a' in the following diagram. The code attempts to find out whether it is possible to connect into this area from the string at D7.

 
   A B C D E F G H J
 9 . . . . a a a . . 9
 8 . . . . a a a a . 8
 7 . . . X O O a a a 7
 6 . . . X X X O a a 6
 5 . . . . + . . a a 5
 4 . . . X . . O a a 4
 3 . . . . X . . a a 3
 2 . . . . . . O a a 2
 1 . . . . . . . . . 1
   A B C D E F G H J

A breakin is found, so the goal is shrunk by removing E9 and J2, then break_in is called again.

In order to see what reading is actually done in order to do this break in, you may load GNU Go in gtp mode, then issue the commands:

 
loadsgf break_in.sgf 
= black

start_sgftrace
= 

break_in D7 E9 F9 G9 E8 F8 G8 H8 G7 H7 J7 H6 J6 H5 J5 H4 J4 H3 J3 H2 J2
= 1 E8

finish_sgftrace vars.sgf
= 

start_sgftrace
= 

break_in D7 F9 G9 F8 G8 H8 G7 H7 J7 H6 J6 H5 J5 H4 J4 H3 J3 H2 J2
= 1 G7

finish_sgftrace vars1.sgf

This will produce two sgf files containing the variations caused by these calls to the breakin code. The second file, `vars1.sgf' will contain quite a few variations.

The break in code makes a list of break ins which are found. When it is finished, the function add_expand_territory_move is called for each break in, adding a move reason.

The break in code is slow, and only changes a few moves by the engine per game. Nevertheless we believe that it contributes substantially to the strength of the program. The break in code is enabled by default in GNU Go 3.6 at level 10, and disabled at level 9. In fact, this is the only difference between levels 9 and 10 in GNU Go 3.6.


13.11 Surrounded Dragons

When is a dragon surrounded?

As has been pointed out by Bruce Wilcox, the geometric lines connecting groups of the opposite color are often important. It is very hard to prevent the escape of this `O' dragon:

 
..........
.....O....
.X.......X
.X...O...X
..........
..........
----------

On the other hand, this dragon is in grave danger:

 
..........
..........
.X.......X
.....O....
.X.......X
.X...O...X
..........
..........
----------

The difference between these two positions is that in the first, the `O' dragon crosses the line connecting the top two `X' stones.

Code in `surround.c' implements a test for when a dragon is surrounded. The idea is to compute the convex hull of the surround set, that is, the set stones belonging to unfriendly neighbor dragons. If the dragon is contained within that hull. If it is, it is said to be surrounded.

In practice this scheme is modified slightly. The implementation uses various algorithms to compute distances and hostile stones are discarded from the surround set when a pair other hostile ones can be found which makes the considered one useless. For example, in the following position the bottom `O' stone would get discarded.

 
O.X.O  
.....
.O.O.
.....
..O..

Also, points are added to the surround set below stones on the second and third lines. This should account for the edge being a natural barrier.

In order to compute distances between corners of the convex hull a sorting by angle algorithm has been implemented. If the distance between a pair enclosing stones is large, the surround status gets decreased to WEAKLY_SURROUNDED, or even 0 for very large ones.

The sorting by angle must be explained. A small diagram will probably help :

 
.O.O.
O...O
..X..
O...O
.O.O.

The sorting algorithm will generate this:

 
.4.5.
3...6
..X..
2...7
.1.8.

That is, the points are sorted by ascending order of the measure of the angle S-G-O, where S is SOUTH, G the (approximated) gravity center of the goal, and O the position of the considered hostile stones.

The necessity of such sorting appears when one tries to measure distances between enclosing stones without sorting them, just by using directly the existing left and right corners arrays. In some positions, the results will be inconsistent. Imagine, for example a position where for instance the points 1,2,3,4,6 and 7 were in the left arrary, leaving only 5 and 8 in the right array. Because of the large distance between 5 and 8, the dragon would have declared weak surrounded or not surrounded at all. Such cases are rare but frequent enough to require the angle sorting.

The following position:

 
O.X.O
.....
.O.O.

This is "more" surrounded than the following position:

 
O.XXXXXX.O
..........
.O......O.

In the second case, the surround status would be lowered to WEAKLY_SURROUNDED.

The surround code is used to modify the escape_route field in the dragon2 data array. When a dragon is WEAKLY_SURROUNDED, the escape_route is divided by 2. If the dragon is SURROUNDED, escape_route is simply set to 0.


13.12 Patterns used by the Influence module

This section explains the details of the pattern databases used for the influence computation.

First, we have the patterns in `influence.db', which get matched symmetrically for both colors.

The patterns in `barriers.db' get matched only for `O' being the player next to move.

The intrusion patterns (`B') are more powerful than the description above might suggest. They can be very helpful in identifying weak shapes (by adding an intrusion source for the opponent where he can break through). A negative inference for this is that a single bad `B' pattern, e. g. one that has a wrong constraint, typically causes 5 to 10 FAILs in the regression test suite.

Influence Patterns can have autohelper constraints as usual. As for the constraint attributes, there are (additionally to the usual ones `O', `o', `X' and `x'), attributes `Y' and `FY'. A pattern marked with `Y' will only be used in the influence computations relevant for the territory valuation, while `FY' patterns only get used in the other influence computations.

The action of an influence pattern is at the moment only used for non-territory patterns as mentioned above, and as a workaround for a problem with `B' patterns in the followup influence.

To see why this workaround is necessary, consider the follwoing situation:

 
..XXX
.a*.O
.X.O.
..XXO

(Imagine that there is `X' territory on the left.)

The move by `O' at `*' has a natural followup move at `a'. So, in the computation of the followup influence for `*', there would be an extra influence source for `O' at `a' which would destroy a lot of black territory on the left. This would give a big followup value, and in effect the move `*' would be treated as sente.

But of course it is gote, since `X' will answer at `a', which both stops the possible intrusion and threatens to capture `*'. This situation is in fact quite common.

Hence we need an additional constraint that can tell when an intrusion pattern can be used in followup influence. This is done by misusing the action line: An additional line

 
>return <condition>;

gets added to the pattern. The condition should be true if the intrusion cannot be stopped in sente. In the above example, the relevant intrusion pattern will have an action line of the form

 
>return (!xplay_attack(a,b));

where `b' refers to the stone at `*'. In fact, almost all followup-specific constraints look similar to this.


13.13 Colored display and debugging of influence

There are various ways to obtain detailed information about the influence computations. Colored diagrams showing influence are possible from a colored xterm or rxvt window.

There are two options controlling when to generate diagrams:

The other options control which diagrams should be generated in these situations. You have to specify at least one of the options above and at least one of the options below to generate any output.

The options below must be combined with one of the two previous ones, or the diagram will not be printed. For example to print the influence diagram, you may combine 0x08 and 0x010, and use the option `-m 0x018'.

Finally, there is the debug option `-d 0x1' which turns on on DEBUG_INFLUENCE. This gives a message for each influence pattern that gets matched. Unfortunately, these are way too many messages making it tedious to navigate the output. However, if you discover an influence source with `-m 0x80' that looks wrong, the debug output can help you to quickly find out the responsible pattern.


13.14 Influence Tuning with view.pike

A useful program in the regression directory is view.pike. To run it, you need Pike, which you may download from http://pike.ida.liu.se/.

The test case `endgame:920' fails in GNU Go 3.6. We will explain how to fix it.

Start by firing up view.pike on testcase endgame:920, e.g. by running pike view.pike endgame:920 in the regression directory.

We see from the first view of move values that filling dame at P15 is valued highest with 0.17 points while the correct move at C4 is valued slightly lower with 0.16. The real problem is of course that C4 is worth a full point and thus should be valued about 1.0.

Now click on C4 to get a list of move reasons and move valuation information. Everything looks okay except that change in territory is 0.00 rather than 1.00 as it ought to be.

We can confirm this by choosing the "delta territory for..." button and again clicking C4. Now B5 should have been marked as one point of change in territory, but it's not.

Next step is to enter the influence debug tool. Press the "influence" button, followed by "black influence, dragons known," and "territory value." This shows the expected territory if black locally moves first everywhere (thus "black influence"). Here we can see that B5 is incorrectly considered as 1.0 points of white territory.

We can compare this with the territory after a white move at C4 (still assuming that black locally moves first everywhere after that) by pressing "after move influence for..." and clicking C4. This looks identical, as expected since delta territory was 0, but here it is correct that B5 is 1.0 points of territory for white.

The most straightforward solution to this problem is to add a non-territory pattern, saying that white can't get territory on B5 if black moves first. The nonterritory patterns are in `barriers.db'.

 
Pattern Nonterritory56

...
X.O
?O.

:8,t

eac
XbO
?Od

;oplay_attack(a,b,c,d,d)

>non_xterritory(e);

In these patterns it's always assumed that `O' moves first and thus it says that `X' can't get territory at B5 (`e' in the pattern). Now we need to be a bit careful however since after `O' plays at `a' and `X' cuts in at `b', it may well happen that `O' needs to defend around `d', allowing `X' to cut at `c', possibly making the nonterritory assumption invalid. It's difficult to do this entirely accurate, but the constraint above is fairly conservative and should guarantee that `a' is safe in most, although not all, cases.


[ << ] [ >> ]           [Top] [Contents] [Index] [ ? ]

This document was generated on November, 27 2004 using texi2html