Ticket #163 (new enhancement)

Opened 1 year ago

Last modified 7 months ago

Semeai upgrades

Reported by: draqo Assigned to: gnugo
Priority: normal Milestone: 3.8
Component: source Version:
Severity: normal Keywords: semeai
Cc: patch: 1

Description

I added some upgrades to semeai module:

1). moves, which fills our dragon eyes are filtered out - this leads to faster analyze and less semeai node limit hits

2). when our dragon has more eyes than opponent, backfilling move should always be tried (previously, ko fight was sometimes chosen instead of backfilling which leads to simple victory)

3). double ko situation is treated as WIN for dragon which has 2 kos adjacent

4). connection pattern 505 is upgraded to give more accurate results - this pattern has big (if not biggest) impact on semeai module and that's why it is in this patch

Attachments

semeai.diff (6.9 kB) - added by draqo on 01/23/07 00:42:18.
fix.diff (420 bytes) - added by draqo on 02/19/07 11:41:02.
Small fix needed after recent change in GnuGo?
upgrade.patch (0.9 kB) - added by draqo on 03/01/07 23:14:15.
Small upgrade - and added another correct move in one test (it showed during work on owl)
fix2.patch (3.6 kB) - added by draqo on 03/07/07 23:31:21.
patch fix - check for a nested ko capture is needed
draqo-semeai.diff (9.0 kB) - added by arend on 04/25/07 01:15:19.
This extracts the semeai parts of all-in-one.zip in #148 and probably supersedes all patches here.
draqo-semeai_b.diff (8.9 kB) - added by arend on 05/03/07 21:24:27.
This is draqo-semeai.diff, somewhat cleaned up
draqo-semeai-tuning.diff (347 bytes) - added by gunnar on 05/09/07 07:56:25.
Rediffed pattern modification from semeai.diff

Change History

01/23/07 00:42:18 changed by draqo

  • attachment semeai.diff added.

01/23/07 00:43:09 changed by draqo

Look for comment from 01/23/07 00:19:53 in ticket 148 for regression description of this ticket.

02/19/07 11:41:02 changed by draqo

  • attachment fix.diff added.

Small fix needed after recent change in GnuGo?

03/01/07 23:14:15 changed by draqo

  • attachment upgrade.patch added.

Small upgrade - and added another correct move in one test (it showed during work on owl)

03/07/07 23:31:21 changed by draqo

  • attachment fix2.patch added.

patch fix - check for a nested ko capture is needed

04/19/07 21:00:42 changed by arend

Could you post a cumulative patch? Generally it is better to post an updated version of your patch, rather than adding a separate "fix patch patch".

04/25/07 01:15:19 changed by arend

  • attachment draqo-semeai.diff added.

This extracts the semeai parts of all-in-one.zip in #148 and probably supersedes all patches here.

04/30/07 20:13:38 changed by arend

The automatic regression tester reports this for drago-semeai.diff as of r2382 (http://trac.gnugo.org/gnugo/regressiontest?details=r2382-pticket-163-draqo-semeai.diff) 

semeai:47       PASS 1 0 R9 [1 0 (PASS|T11|T9|R9|T10|R10|R11)]
trevord:680     FAIL R15 [S16]
nngs4:270       PASS M1 [!P5]
9x9:120         PASS F1 [F1]
9x9:650         PASS E1 [E1]
4 PASS
1 FAIL
Total nodes: 1709731659 3334380 12425491 (+0.018% -0.17% -0.04%)

I looked through these, the breakage looks very good.

05/03/07 21:24:27 changed by arend

  • attachment draqo-semeai_b.diff added.

This is draqo-semeai.diff, somewhat cleaned up

05/09/07 07:56:25 changed by gunnar

  • attachment draqo-semeai-tuning.diff added.

Rediffed pattern modification from semeai.diff

05/09/07 22:45:41 changed by gunnar

The CC505 pattern (draqo-semeai-tuning.diff) is not doing well on its own, 

semeai:126      FAIL 2 2 T15 [2 2 (P19|S17|T17)]
semeai:129      FAIL 1 1 S17 [3 3 (Q19|S17|T17)]
semeai:141      FAIL 1 1 F2 [2 2 (F1|F2|F3)]
semeai:143      FAIL 1 1 A2 [3 3 (F1|F3)]
owl1:298        FAIL 0 [2 Q3]
owl1:299        FAIL 1 PASS [1 Q1]
nngs2:540       PASS B8 [B8|B9]
nngs3:520       PASS T8 [!N3]
13x13c:43       FAIL K12 [G13]
gifu03:507      pass
9x9:650         fail
nngs3:490       fail
3 PASS (2 PASS, 1 pass)
9 FAIL (7 FAIL, 2 fail)
Total nodes: 1709187088 3323727 12427144 (-0.11% -0.4% -0.0024%)

The semeai and owl1 failures seem ko related and probably CC505 needs to be combined with the patches in #164.

The 9x9:650 and nngs3:490 failures are apparently interference with the changeset [2382], ticket #75.

10/24/07 21:16:43 changed by gunnar

  • milestone changed from 3.7.11 to 3.8.