Re: [gnugo-devel] auto generated regression tests

[Gunnar Farneback]

Dan wrote:
> As a general comment, the dragon_status is not a very important
> field.

The status field of the dragon_data struct is not very accurate no,
but that is not what the gtp dragon_status function reports. The
purpose of the function is to return the engine's best estimate of a
dragon's status, whatever internal data structures are involved. The
function looks like this:

| /* Function:  Determine status of a dragon.
|  * Arguments: vertex
|  * Fails:     invalid vertex, empty vertex
|  * Returns:   status ("alive", "critical", "dead", or "unknown"),
|  *            attack point, defense point. Points of attack and
|  *            defense are only given if the status is critical and the
|  *            owl code is enabled.
|  *
|  * FIXME: Should be able to distinguish between life in seki
|  *        and life with territory. Should also be able to identify ko.
|  */
| 
| static int
| gtp_dragon_status(char *s, int id)
| {
|   int i, j;
| 
|   if (!gtp_decode_coord(s, &i, &j))
|     return gtp_failure(id, "invalid coordinate");
| 
|   if (BOARD(i, j) == EMPTY)
|     return gtp_failure(id, "vertex must not be empty");
| 
|   silent_examine_position(BLACK, EXAMINE_DRAGONS);
|   
|   /* FIXME: We should also call the semeai module. */
| 
|   if (dragon[POS(i, j)].owl_status == UNCHECKED) {
|     if (dragon[POS(i, j)].status == ALIVE)
|       return gtp_success(id, "alive");
|   
|     if (dragon[POS(i, j)].status == DEAD)
|       return gtp_success(id, "dead");
|   
|     if (dragon[POS(i, j)].status == UNKNOWN)
|       return gtp_success(id, "unknown");
| 
|     assert(dragon[POS(i, j)].status == CRITICAL); /* Only remaining 
possibility. */
|     return gtp_success(id, "critical");
|   }
| 
|   /* Owl code active. */
|   if (dragon[POS(i, j)].owl_status == ALIVE)
|     return gtp_success(id, "alive");
|   
|   if (dragon[POS(i, j)].owl_status == DEAD)
|     return gtp_success(id, "dead");
|   
|   if (dragon[POS(i, j)].owl_status == UNKNOWN)
|     return gtp_success(id, "unknown");
|   
|   assert(dragon[POS(i, j)].owl_status == CRITICAL);
|   /* Status critical, need to return attack and defense point as well. */
|   gtp_printid(id, GTP_SUCCESS);
|   gtp_printf("critical ");
|   gtp_print_vertex(I(dragon[POS(i, j)].owl_attack_point),
|                    J(dragon[POS(i, j)].owl_attack_point));
|   gtp_printf(" ");
|   gtp_print_vertex(I(dragon[POS(i, j)].owl_defense_point),
|                    J(dragon[POS(i, j)].owl_defense_point));
|   return gtp_finish_response();
| }

This should probably be rewritten today using the matcher_status field
of the dragon_data struct. (The function predates the matcher_status
field and, as can be deduced from a comment, was written while the owl
code was still an experimental option.)

> Probably these tests should be in terms of either matcher_status or
> owl tests.

No, dragon_status is perfectly fine. Owl tests are of course fine as
well, but only really needed when ko results are involved. See
ld_owl.tst for lots of examples.

/Gunnar