Issue 90 19 June, 2004
_______________________________________________________________________________
Core Warrior is a newsletter promoting the game of Corewar. Emphasis is
placed on the most active hills - currently the '94 no-pspace, '94 draft
and beginner hills. Coverage will follow wherever the action is. If you
haven't a clue what I'm talking about then check out these five-star
Internet locals for more information:
FAQs are available from:
http://www.koth.org/corewar-faq.html
http://homepages.paradise.net.nz/~anton/cw/corewar-faq.html
Web pages are at:
http://www.koth.org/ ;KOTH
http://sal.math.ualberta.ca/ ;Sal Hills
http://para.inria.fr/~doligez/corewar ;Planar
http://www.ociw.edu/~birk/corewar ;C.Birk
http://www.corewar.info/ ;Fizmo
http://www.corewar.co.uk/ ;J.Metcalf
Newbies should check the above pages for the FAQs, language specification,
guides, and tutorials. Post questions to rec.games.corewar. All new players
are infinitely welcome!
_______________________________________________________________________________
Greetings...
Yes, I am back!!!
There have been a lot of issues published since I prepared my last Core
Warrior. It's exactly 30 issues since I sent the first one to r.g.c. ;-)
You are wondering who am I? This time it's not John Metcalf. This time
Core Warrior has been prepared by me, Christian Schmidt.
And there is one thing I really wanted to say for a very long time:
_ _ _
| | | | | |
| |_| |__ __ _ _ __ | | __ _ _ ___ _ _
| __| '_ \ / _` | '_ \| |/ / | | | |/ _ \| | | |
| |_| | | | (_| | | | | < | |_| | (_) | |_| |
\__|_| |_|\__,_|_| |_|_|\_\ \__, |\___/ \__,_|
__/ |
|___/
___ _
|_ | | |
| | ___ | |__ _ __
| |/ _ \| '_ \| '_ \
/\__/ / (_) | | | | | | |
\____/ \___/|_| |_|_| |_|
for providing us for 5 YEARS continuously with new issues of the
Core Warrior.
Since #70 he has prepared all issues of Core Warrior for us. I think that
is a good reason beside many, many others to admit him into the Corewar
Hall of Fame!!!
---
I think I should remind everyone again that there will be a corewar meeting
on July 23-25, 2004 in Berlin.
All player who would like to participate in the meeting should contact me as
soon as possible. I will pick all people up at the airport, train station,
etc., help find in advance accommodation (hotel etc.) and take care of them
during their stay in Berlin.
The following activities are planned so far:
- Sightseeing tour
- Berlin by night
- A pen'n'paper corewar tournament (no PC allowed for writing the warriors ;-)
- Corewar Workshops (who would also like to give a workshop? I hope I am not
the only one!!!)
Further ideas are welcome. I'll keep the website updated with the latest news.
---
Redcode Frenzy News: The results of the Capture The Flag '88 Round are online
now. Congratulations to the winner: Chip Wendell.
Round 19 will be organised by me (Christian Schmidt) and Round 20 by Sascha
Zapf.
---
Web-News: John Metcalf's homepage has moved to:
http://www.corewar.co.uk/
Terry Newton's page has move to:
http://www.infionline.net/~wtnewton/corewar/
---
Okay, coming to this issue. It has turned out much larger than expected.
This time, we introduce the interesting LP hill on SAL including some hints
and ExtraExtras. But that's not all. We have some more hints and ExtraExtras
for you, which shows again after so many years of corewar, there are many
things waiting to discover :-)
I hope we are able to push your interests to a new level and will see you
soon on one of the hills. The battle between the corewar warriors still
goes on.
Now enjoy this issue!!!!
-- Christian Schmidt
'94 No Pspace Hill
*******************
______________________________________________________________________________
Current Status of the KOTH.ORG '94 No Pspace Hill:
# %W/ %L/ %T Name Author Score Age
1 33/ 27/ 39 Yorba Roy 139 65
2 30/ 23/ 47 paper(paper(paper(clear)) Sascha Zapf 138 35
3 42/ 47/ 11 Arrow Christian Schmidt 138 134
4 41/ 44/ 15 Dandelion 3 Christian Schmidt 138 52
5 42/ 46/ 12 O--* Bremer/Schmidt 137 60
6 26/ 16/ 57 Maelstrom Roy van Rijn 137 75
7 41/ 45/ 14 HazyLazy C 11 CS SG 136 143
8 31/ 25/ 44 Triturus Roy van Rijn 136 19
9 31/ 27/ 42 Gremlin van Rijn/Grabun 136 165
10 41/ 47/ 12 Bottomless Pit Fizmo/Roy 136 96
11 31/ 27/ 42 slime test 1.00 David Houston 135 261
12 32/ 30/ 38 Frantic Roy van Rijn 134 48
13 41/ 48/ 11 Speeed Christian Schmidt 133 2
14 40/ 47/ 13 High Voltage Sascha Zapf 133 9
15 27/ 21/ 53 Black Knight Christian Schmidt 133 14
16 31/ 29/ 40 Son of Vain Oversby/Pihlaja 133 2499
17 27/ 22/ 51 Bvowkium Fizmoxid bvowk/fizmo 131 1
18 38/ 45/ 16 Solar Flare test John Morahan 131 10
19 41/ 51/ 8 Aoshi A 34 Steve Gunnell 131 37
20 31/ 32/ 37 Spiritual Black Dimension Christian Schmidt 130 290
The hill has aged just 94 since last issue, bringing it to a total age of 3253.
All but seven warriors have been pushed from the hill.
Those who didn't last the duration include Ironic Imps (age 403), Armadillo
(233), Quirk (207), Arctica (178), Devilstick (167) and Kenshin D 10 (111).
Warriors which entered the hill since last issue and were pushed off age
50+: something strange (87) and Deathstar (86).
Koth report: 7 warriors resided at number one for at least 5 challenges.
Most often seen there has been Maelstrom (24), followed by Yorba (16),
Bottomless Pit (11), Arrow (8), HazyLazy C 11 (5), O--* (5) and Son of
Vain (5).
Son of Vain again increased it's oldest Koth record, taking first place
age 2431.
_______________________________________________________________________________
The '94 No Pspace Hall of Fame: * indicates the warrior is still active.
Pos Name Author Age Strategy
1 Son of Vain Oversby/Pihlaja 2499 * Q^4 -> Stone/imp
2 Blacken Ian Oversby 1363 Q^2 -> Stone/imp
3 Reepicheep Grabun/Metcalf 1347 Q^4 -> Paper/stone
4 nPaper II Paul-V Khuong 1270 MiniQ^3 -> Paper
5 Uninvited John Metcalf 1130 MiniQ^3 -> Stone/imp
6 Hazy Test 63 Steve Gunnell 1119 Scanner
7 Behemot Michal Janeczek 1078 MiniQ^3 -> Bomber
8 Olivia Ben Ford 886 Q^4 -> Stone/imp
9 Recon 2 David Moore 826 Scanner
10 Keyser Soze Anton Marsden 823 Qscan -> Bomber/paper/imp
11 Quicksilver Michal Janeczek 789 Q^4 -> Stone/imp
12 Eraser II Ken Espiritu 781 Scanner
13 Inky Ian Oversby 736 Q^4 -> Paper/stone
14 Toxic Spirit Philip Thorne 685 One-shot
15 Jinx Christian Schmidt 662 Q^3 -> Scanner
16 Blade Fizmo 643 Qscan -> Scanner
17 Return of Vanquisher Lukasz Grabun 640 Q^4 -> Bomber
18 Revenge of the Papers Fizmo/Roy 605 Q^4 -> Paper
19 Jade Ben Ford 600 Q^4 -> Stone/imp
20 Firestorm John Metcalf 589 MiniQ^3 -> Paper/imp
21 Claw Fizmo 525 Qscan -> Scanner
22 G3-b David Moore 503 Twoshot
23 Thunderstrike Lukasz Grabun 484 Q^4 -> Stone/imp
24 Vanquisher Lukasz Grabun 469 Q^4 -> Bomber
25 Revival Fire P.Kline 468 Bomber
26 The Phantom Menace Anton Marsden 465 Qscan -> Paper/imp
27 The Stormkeeper Christian Schmidt 460 Q^3 -> Stone/imp
28 Positive Knife Ken Espiritu 449 Q^4 -> Stone/imp
29 Boys are Back in Town Philip Kendall 441 Scanner
= Zooom... John Metcalf 441 Scanner
31 G2-b David Moore 413 Twoshot
32 Ironic Imps Roy van Rijn 403 Q^4 -> Stone/imp
33 Qtest Christian Schmidt 394 Q^3 -> Paper
34 Stalker P.Kline 393 Scanner
35 Decoy Signal Ben Ford 378 Q^4 -> Paper/imp
36 unheard-of Christian Schmidt 355 Q^4 -> Paper/imp
37 Hazy Lazy ... again Steve Gunnell 350 Scanner
= Dawn Roy van Rijn 350 Q^4 -> Paper/imp
39 Static Miz 344 Q^4 -> Stone/imp
40 devilish 2 David Houston 330 Q^4 -> Stone/imp
= Vain Ian Oversby 330 Q^2 -> Stone/imp
42 Omnibus John Metcalf 327 Q^2 -> Stone/imp
43 Win! David Moore 322 Scanner
44 Return of the PendragonChristian Schmidt 318 Q^4 -> Stone/imp
45 Numb Roy van Rijn 312 Q^4 -> Paper/stone
46 Hazy Lazy... Steve Gunnell 309 Scanner
47 Now you're in trouble! John Metcalf 290 Q^4 -> Stone/imp
= Spiritual Black Dim. Christian Schmidt 290 * Q^4.5 -> Paper/stone
49 Hazy Lazy ... reborn Steve Gunnell 284 Scanner
50 Candy II Lukasz Grabun 282 Q^4 -> Stone/imp
= Pixie Lukasz Grabun 282 Q^4 -> Stone/imp
There are only 2 active warriors in this extended HoF, compared to last
issue's 2. The warriors are represented by 18 Stone/imps, 13 scanners,
6 paper/imps, 4 bombers and paper/stones and 3 each of one-shots
and papers.
18 authors hold a place in the HoF. Schmidt has most, with 9 warriors.
Grabun and Metcalf both have 6 - Gunnell, Moore, Oversby and van Rijn
4 each - Ford has 3 - Espiritu, Janeczek, Kline and Marsden have 2 each
and the remaining 6 authors, just the one warrior.
Of these redcoders, 3 have 4 different strategies in the HoF. They are
Metcalf, van Rijn and Schmidt.
Next warrior which will enter is David Houston's slime test 1.00.
'94 Draft Hill
*******************
_______________________________________________________________________________
Current Status of the KOTH.ORG '94 Draft Hill:
# %W/ %L/ %T Name Author Score Age
1 43/ 43/ 14 Lord of the Core Christian Schmidt 142 2
2 35/ 29/ 37 Spiritual Black Dimension Christian Schmidt 141 67
3 41/ 42/ 17 Cyberpunk Christian Schmidt 140 75
4 43/ 47/ 10 Arrow Christian Schmidt 139 48
5 35/ 34/ 31 Sunrise Zul Nadzri 136 77
6 32/ 28/ 41 Devilish 2 David Houston 136 37
7 31/ 26/ 43 Reepicheep Grabun/Metcalf 135 497
8 26/ 17/ 56 Maelstrom Roy van Rijn 135 27
9 27/ 18/ 55 Black Knight Christian Schmidt 135 9
10 31/ 27/ 42 Armadillo Lukasz Grabun 134 8
11 33/ 32/ 35 KL Zul Nadzri 134 7
12 37/ 40/ 23 Dettol Test 45 Steve Gunnell 134 3
13 29/ 26/ 45 Gremlin van Rijn/Grabun 133 53
14 40/ 48/ 12 Dandelion 3 Schmidt/Zapf 133 51
15 27/ 22/ 51 Muscle Hammer Zul Nadzri 132 32
16 26/ 20/ 54 Bvowkium Fizmoxid bvowk/fizmo 132 20
17 29/ 27/ 43 season of adventure John Metcalf 131 6
18 31/ 31/ 38 Arctica Christian Schmidt 131 57
19 38/ 45/ 17 something strange John Metcalf 130 33
20 39/ 49/ 12 Speeed Christian Schmidt 130 1
After an age of 32, 11 warriors remain on the hill from last issue.
Those who didn't last the duration include Son of Vain (441), Return of
Vanquisher PsP (284) and Bitter Sweet (262) leaving the hill with just
one warrior at an age higher than 100.
Reepicheep is now by far the oldest warrior with an age of 497.
p-Warriors are in top spots now but pure scanners are scoring surprisingly
well and can still survive for a lengthy term on the hill.
_______________________________________________________________________________
The '94 Draft Hall of Fame: * indicates the warrior is still active.
Pos Name Author Age Strategy
1 Reepicheep Grabun/Metcalf 497 * Q^4 -> Paper/stone
2 Son of Vain Oversby/Pihlaja 441 Q^4 -> Stone/imp
3 Return of Vanquisher Lukasz Grabun 284 Q^4 -> Bomber
4 Bitter Sweet Lukasz Grabun 262 Q^4 -> Stone/imp
5 CrazyShot 2 Christian Schmidt 249 Q^4 -> One-shot
6 Herbal Avenger Michal Janeczek 237 Scanner
7 Bustling Spirit Christian Schmidt 216 P-warrior
8 Revenge of the Papers Fizmo/Roy 204 Q^4 -> Paper
9 Uninvited John Metcalf 194 MiniQ^3 -> Stone/imp
10 Blowrag Metcalf/Schmidt 192 Q^4 -> Paper/imp
11 Incredible! John Metcalf 180 Paper/imp
12 Wallpaper Christian Schmidt 175 Q^4 -> Paper/stone
13 Mantrap Arcade Dave Hillis 170 P-warrior
14 Numb Roy van Rijn 167 Q^4 -> Paper/stone
15 PolyPap Jakub Kozisek 160 Qscan -> Paper
16 Recon 2 David Moore 156 Scanner
17 Joyful Maw Dave Hillis 143 P-warrior
18 Paperazor Christian Schmidt 141 Q^4 -> Paper
19 Sunset David Moore 138 P-warrior
20 Dawn Roy van Rijn 137 Q^4 -> Paper/imp
21 Self-Modifying Code Ben Ford 132 P-warrior
22 Combatra David Moore 131 Boot distance calculator
23 unheard-of Christian Schmidt 128 Q^4 -> Paper/imp
24 Mad Christian Schmidt 123 P-warrior
25 Microvenator Michal Janeczek 122 P-warrior
No new entries this time. Reepicheep is the only active warrior. Next to
enter will be Sunrise and Cyberpunk.
SAL Beginners Hill
*******************
_______________________________________________________________________________
Current Status of the SAL Beginners Hill:
# %W/ %L/ %T Name Author Score Age
1 50.7/ 30.9/ 18.4 Studie strikes back Sascha Zapf 170.6 29
2 49.4/ 31.3/ 19.3 Shooting Star v0.262 brx 167.4 7
3 50.1/ 35.1/ 14.8 Frullato LAchi 165.2 46
4 48.7/ 33.7/ 17.6 ssoBeht Sascha Zapf 163.8 20
5 49.1/ 42.2/ 8.8 Star Sascha Zapf 156.0 34
6 33.9/ 20.7/ 45.4 Uneven Twins Sascha Zapf 147.1 33
7 43.4/ 41.1/ 15.5 French Kiss LAchi 145.8 27
8 30.3/ 18.9/ 50.8 Veenah V0.8 Sascha Zapf 141.6 5
9 42.1/ 45.8/ 12.0 scizzory Miz 138.5 76
10 40.7/ 43.0/ 16.3 Neverland II John Metcalf 138.3 75
11 27.6/ 19.1/ 53.4 Paper of yesterday Sascha Zapf 136.1 81
12 26.1/ 17.5/ 56.4 Tom David Moore 134.7 17
13 28.7/ 26.4/ 45.0 Fatty boy G.Labarga 131.0 87
14 37.2/ 43.5/ 19.3 Lifer Sascha Zapf 130.9 4
15 27.1/ 31.4/ 41.5 Manufactured paper Neo 122.9 3
16 21.7/ 21.3/ 57.0 Grand Jury Andrew Hunter 122.1 1
17 17.2/ 17.8/ 65.0 Unknown Anonymous 116.6 26
18 18.4/ 20.5/ 61.1 Enigma brx 116.4 13
19 16.8/ 18.4/ 64.9 nGrave Ken Hubbard 115.2 38
20 19.9/ 24.9/ 55.1 Smaug Imp Ken Hubbard 114.9 43
21 17.4/ 20.1/ 62.5 Plastered Andrew Hunter 114.7 2
22 18.9/ 26.3/ 54.9 test3 miz 111.5 90
23 29.8/ 51.3/ 18.8 Stoney Miz 108.3 77
24 27.1/ 58.5/ 14.4 Sugartooth Miz 95.8 41
25 24.9/ 57.9/ 17.2 Brimstone John Metcalf 91.8 23
Just 4 challenges, so not much action here :-/ Is the hill too strong for the
beginner?
'94 LP Hill
*******************
We will try to focus in every coming issue on a different hill. This time we
have chosen SAL's Limited Process Hill.
The hill specifications are:
coresize: 8000
max. processes: 8
duration: after 80,000 cycles, a tie is declared.
max. entry length: 200
minimum distance: 200
rounds fought: 250
instruction set: ICWS '94 Draft
_______________________________________________________________________________
Current Status of the SAL Limited Process Hill:
# %W/ %L/ %T Name Author Score Age
1 49.6/ 21.8/ 28.6 The King Neo 177.3 8
2 40.8/ 30.8/ 28.3 Lord of the Core Christian Schmidt 150.9 2
3 38.0/ 27.3/ 34.7 Duke Christian Schmidt 148.6 26
4 40.0/ 33.1/ 26.9 Reclon 2 G.Labarga 146.8 9
5 40.3/ 36.1/ 23.5 Viper's Bugloss Christian Schmidt 144.5 27
6 38.6/ 33.9/ 27.5 Muffineer Christian Schmidt 143.4 21
7 29.2/ 16.1/ 54.7 Eve 15 bvowk + Fizmo 142.3 55
8 29.3/ 23.1/ 47.6 The LP Machine Zul Nadzri 135.5 5
9 38.4/ 42.6/ 18.9 fbomber Michal Janeczek 134.2 107
10 29.3/ 24.9/ 45.9 Roskilde Christian Schmidt 133.6 86
11 24.6/ 17.1/ 58.4 8366-4510-cs-sdk-eve15 bvowk 132.1 57
12 29.0/ 26.2/ 44.8 xXx Zul Nadzri 131.9 6
13 33.2/ 34.7/ 32.2 Achillu 2004 - Revenge LAchi+Fizmo+Neogry 131.7 1
14 34.8/ 38.1/ 27.2 Revenge of the Silklander Christian Schmidt 131.4 65
15 39.0/ 48.0/ 13.0 FlipSide Zul Nadzri 129.9 68
16 36.8/ 43.9/ 19.3 ~~ The 9th Process ~~ Zul Nadzri 129.8 98
17 30.3/ 31.2/ 38.5 Histeric imps G.Labarga 129.4 13
18 38.3/ 47.4/ 14.2 Asfixia G.Labarga 129.2 12
19 36.1/ 44.1/ 19.8 100% Evil Roy van Rijn 128.0 46
20 29.7/ 32.2/ 38.1 Royal guard G.Labarga 127.3 14
21 34.8/ 42.5/ 22.7 Mad LP Christian Schmidt 127.2 74
22 32.9/ 39.8/ 27.3 Dental Disease Christian Schmidt 126.1 78
23 30.9/ 36.2/ 32.9 wandering redcoder John Metcalf 125.7 4
24 25.5/ 25.5/ 49.0 4310-4309-xt430-0-eve15 bvowk 125.5 25
25 33.6/ 42.2/ 24.2 scan G.Labarga 125.1 24
The total age of this new hill is 118. Two warrior reached a three-digit age
while the oldest of the two, The 8 Kings (112), has been recently pushed off.
The hill is dominated by Neo's The King which has a 26 points gap to the 2nd
place. He takes a strong advantage by handshaking with several other of his
warriors.
Bvowk's evolved warrior 8366-4510-cs-sdk-eve15 is still going strong as well
as its human-optimized brother Eve 15.
Do we see soon the first purely evolved warrior which passes the requirements
for a hill's Hall of Fame? We will be carefully watching this....
_______________________________________________________________________________
The Story - A Short History of '94 LP Hills by Christian Schmidt
The Wasted Youth hill
---------------------
The beginnings of the first '94 LP hill lay in the ancient ages around 1996
and only few things are known from this time. John K. Lewis opened the hill.
The first of the LP hills closed in the beginning of '97.
The Pizza LP hill
-----------------
Beginning March 1997 some threads on r.g.cw discussed new hills with different
specs. Anton Marsden mentioned in that sense on r.g.cw:
"LP hills provide some interesting new strategies:
- papers which are constructed differently to the "standard" '88/'94 papers.
- self-repairing programs
- multiple function programs, eg. 4 stones combined with 4 scanners
Add pspace to this and it could be a great hill."
Anton Marsden, March 2, 1997
After an e-mail vote the Pizza LP hill was announced by Thomas "Thos" Davies
on March 14, 1997 using the same specification as the Wasted Youth Hill.
Ian Sutton's Limb Rending Incisors was dominating this hill.
Around spring 2001 the Pizza Hills closed.
SAL LP hill
-----------
End of 2003 Joonas Pihlaja and Barkley Vowk announced the SAL server. It
provides beside others, a new LP hill following the spec of the previous two.
It's quite a young hill and recently reached the age of 100. It has received
since the beginning a continuously mediocre traffic.
-- Christian Schmidt
_______________________________________________________________________________
The Hint - a few remarks about LP strategies by Christian Schmidt
If only 8 processes are available the process limit can be reached very fast.
If a warrior is running at its process limit, a spl acts just like a nop.
This leads to some very specific differences to the 94nop and 94draft hill:
- Spl-bombs are by far less effective than jmp #0 bombs
- low-process papers are scoring better than silk-style
- different multiple strategy warriors are effective. For example, 2 stones
with one process each and the rest in imps or a number of parallel running
stones and scanners, etc.
Check the following warriors at Koenigstuhl's Infinite LP Hill to get an
impression of LP warrior design:
Paper: Revenge of the Silklander, Grained Paper, Achillu 94, Elusive, Oldtimer,
Skinner, 2LPap
Stone/Paper: Roskilde, Evil Twin Brother, Tempest
Scanner, Oneshots: lp-Shot v0.1, LP-Scan v0.1, One Hop
Stone/Imps: Philosopher's Stone
Stones, Vamps: fbomber, Limb Rending Incisors, The 8 Kings
-- Christian Schmidt
_______________________________________________________________________________
Extra Extra - New approaches in evolving warriors and finding new strategies
by Christian Schmidt and Barkley Vowk
The warrior 189602-1978-xt642-2-eve15 shows a really interesting scoring
pattern on the LP hill. Although it is a very defensive strategy it's able to
beat most of the papers quite nicely.
A closer look at the code shows, that the warrior executes just 8 instructions.
The remaining 192 lines are not in use and seems to act as a huge decoy only.
The 'active' part of the warrior containing some additional comments is shown
below:
mov.i $ 7 , { 2 ;boot the imp
spl.a $ 5 , # 1867 ;split behind the imp-gate
spl.ab $ 2317 , $ 931 ;launch the booted imp
djn.a # -1 , < -145 ;imp-gate / djn-train
djn.f # -1 , < -145 ;\
jmp.ba # 21 , < -43 ; ->safety instructions?
mov.b # 2667 , $ 8 ;useless instruction
mov.i # -1 , $ 1 ;the imp instruction
The first impression is: What? That's all!?! Just two imps and an imp-gate?
But how can it be such defensive and be able to beat LP papers?
Only on a second view it unfurls its barbaric brilliancy: A strategy which was
never seen in a hand-coded warrior for the LP-hill before.
Both imps are running fast enough to being not caught by a coreclear or a
scanner wipe on their way through the core. And if they overwrites an opponent
they convert it into an imp. And they are fast enough to catch the papers
while still in process of replicating themselves. At the end of all this waits
the imp-gate killing every imp that approaches. This means if the two imps
were able to convert all processes of the paper into imps while running
through the core the paper will lose.
The next step after understanding the strategy was to write a hand-coded
version and finally optimize it. For some additional points a quickscanner
was added which works also as a decoy.
And without further ado, here is the final code:
;redcode-lp
;name Eve 15
;strategy No humans were used in the creation
;strategy of this strategy
;author bvowk + Fizmo
;assert 1
;------>qscan constants
zero equ qbomb
qtab3 equ qbomb
qc2 equ ((1+(qtab3-qptr)*qy) % CORESIZE)
qb1 equ ((1+(qtab2-1-qptr)*qy) % CORESIZE)
qb2 equ ((1+(qtab2-qptr)*qy) % CORESIZE)
qb3 equ ((1+(qtab2+1-qptr)*qy) % CORESIZE)
qa1 equ ((1+(qtab1-1-qptr)*qy) % CORESIZE)
qa2 equ ((1+(qtab1-qptr)*qy) % CORESIZE)
qz equ 2108
qy equ 243
;------>qbomb constants
qoff equ -87
qstep equ -7
qtime equ 14
;------>eve 15 constants
iAwa equ 4174
train equ 7903
;------>eve 15 code
pGo mov.i imp, *2
spl imp
spl iAwa
djn.a #0, <train
djn.f #0, <train
jmp.ba #0, <train
imp mov.i #-1, 1
for 51
dat 0, 0
rof
;------>qscan code
qbomb dat >qoff, >qc2
dat 0, 0
dat 0, <qb1
qtab2 dat 0, <qb2
dat 0, <qb3
for 16
dat 0, 0
rof
dat zero - 1, qa1
qtab1 dat zero - 1, qa2
for 42
dat 0, 0
rof
qgo sne qptr+qz*qc2, qptr+qz*qc2+qb2
seq <qtab3, qptr+qz*(qc2-1)+qb2
jmp q0, }q0
sne qptr+qz*qa2, qptr+qz*qa2+qb2
seq <qtab1, qptr+qz*(qa2-1)+qb2
jmp q0, {q0
sne qptr+qz*qa1, qptr+qz*qa1+qb2
seq <(qtab1-1), qptr+qz*(qa1-1)+qb2
djn.a q0, {q0
sne qptr+qz*qb3, qptr+qz*qb3+qb3
seq <(qtab2+1), qptr+qz*(qb3-1)+(qb3-1)
jmp q0, }q1
sne qptr+qz*qb1, qptr+qz*qb1+qb1
seq <(qtab2-1), qptr+qz*(qb1-1)+(qb1-1)
jmp q0, {q1
sne qptr+qz*qb2, qptr+qz*qb2+qb2
seq <qtab2, qptr+qz*(qb2-1)+(qb2-1)
jmp q0
seq >qptr, qptr+qz+(qb2-1)
jmp q2, <qptr
seq qptr+(qz+1)*(qc2-1),qptr+(qz+1)*(qc2-1)+(qb2-1)
jmp q0, }q0
seq qptr+(qz+1)*(qa2-1),qptr+(qz+1)*(qa2-1)+(qb2-1)
jmp q0, {q0
seq qptr+(qz+1)*(qa1-1),qptr+(qz+1)*(qa1-1)+(qb2-1)
djn.a q0, {q0
jmz.f pGo, qptr+(qz+1)*(qb2-1)+(qb2-1)
;------>qbomb code
q0 mul.b *2, qptr
q2 sne {qtab1, @qptr
q1 add.b qtab2, qptr
mov qtab3, @qptr
qptr mov qbomb, }qz
sub #qstep, qptr
djn -3, #qtime
jmp pGo
end qgo
And now a few words on how this warrior was evolved:
Each run of the evolver starts out with "seeding the pool", this is done by
running a utility called makepop. makepop creates a new warrior by filling it
full of code chosen randomly from instruction distributions pulled from
successful handcoded warriors, and then checking to see if it can pass test
battle against a random warrior from a benchmark set.
Once the pool is filled with warriors, I run the server program master. Master
reads all the generated warriors, the benchmark warriors, then shuffles the
generated warriors into species. The evolver considers a species to be a group
of warriors in the pool that breed only with other members of their group. The
benchmark typically sits around 6-12 warriors. When master is online and
listening, I start clients on any computers I have available. The clients
contact the master, and checkout a section of the pool to calculate scores
for. Once the client finishes calculating the initial warrior scores, it
returns the information to the server and checks out another small section.
Once all the warriors have an initial score assigned, the server begins
sending the clients the completed pool. The clients breed new warriors. Any
new warriors that exceed the score of the current lowest warrior in a species
are copied in over that warrior, removing it from the pool.
The "breeding" process selects two warriors of the same species, one of the
parents is chosen to be the source. After each instruction is copied to the
child warrior, there is a chance the source will be swapped to a different
parent warrior, and instructions will be copied from that parent instead.
On average there are 4 cross-overs between parents in each child. As well,
there is a chance of mutation for every instruction, this can take the form
of either a small change to the current instruction, dropping this instruction
completely, or adding a new instruction. Once the child has been filled with
code, it is validated against a random benchmark warrior to ensure it is worth
the cycles to calculate a full score. If it passes, we continue as above, if
the child scores better than the current low, it replaces the lowest warrior
in the pool.
After the client has found enough improved warriors, it returns its pool to the
server, which adds the improved warriors to the main pool. This will continue
until someone stops the master, and the clients exit when they can check-out no
more work.
There is also a simple proxy server, to allow clients to be run on machines that
don't have direct access to the internet. The proxy checks out work (from either
the master, or another proxy), and hands it out to clients that connect to it.
Once the proxy recieves enough improved warriors, it checks them back into the
main server, and requests a new pool.
-------- ---------
|Master|---------|Clients|
-------- | ---------
|
| --------- --------------
------|Proxy S|------|More Clients|
--------- --------------
|
|
| --------- ------------------
------|Proxy S|----|Yet More Clients|
--------- ------------------
I have found that best results are obtained by starting the validation score
very low, and ramping it up as the baseline score of the warriors improves.
When most warriors are capable of scoring >50% of the tie barrier (the score
achieved with 100% ties), the validation score is bumped so only warriors that
can exceed the tie barrier will have their score calculated. This prevents
warriors that have no strong offensive ability from breeding out all the very
offensive warriors from the pool, as offensive seem to have a more difficult
time finding improvements. The mutation rate is started quite high, about 60%
chance per instruction, and is dropped quickly as the warriors improve. Towards
the end, the mutation rate is typically held somewhere around 2% chance per
instruction. If the mutation rate is not dropped, the mutation rate will cause
the pool to flail around aimlessly and never converge. If the mutation rate is
started too low, there will not be enough variation in the pool, and it is
unlikely to find optimizations. The typical path of the evolver is to start
with low scores, quickly increase as we approach the tie barrier, then there is
a significant slowdown after the validation score bump described above. Once
most of the population starts to shift towards more offensive warriors, the
climb upwards starts again. This will continue until the pool converges on a
monoculture, where all warriors effectively use the same tactics, and we no
longer find further improvements. At this point, the pool can be considered
done.
I generally run the evolver several times from a random start to until the pool
converges on a monoculture, then collect the top warriors from the finished
pools into a new pool, and re-run the evolver hoping for further improvements.
This strategy has proven quite successful in the past. Typically it will take
days or weeks to run each pool to completion.
As mentioned above, the code generated by the evolver is seeded from patterns
seen in human coded warriors. The theory here is to emulate natural organisms.
In nature there are only a limited number of working combinations of proteins,
and the creation of un-workable proteins is suppressed through various means.
The same is true in corewars, wherein we want to maximize the working
combinations we see in "nature" (human coded warriors), and supress the random
ones that we don't see. To do this, I devised a system I think of as chains.
A chain might look like:
Start instruction:
SEQ occured 69 times (39.884%)
MOV occured 49 times (28.324%)
SNE occured 23 times (13.295%)
SPL occured 17 times (9.827%)
ADD occured 7 times (4.046%)
SUB occured 3 times (1.734%)
JMZ occured 3 times (1.734%)
DIV occured 1 times (0.578%)
NOP occured 1 times (0.578%)
2nd Instruction:
If previous instruction was spl:
SEQ occured 771 times (60.613%)
SNE occured 170 times (13.365%)
DAT occured 137 times (10.770%)
SPL occured 69 times (5.425%)
JMP occured 36 times (2.830%)
MOV occured 34 times (2.673%)
ADD occured 16 times (1.258%)
DJN occured 6 times (0.472%)
SUB occured 3 times (0.236%)
JMZ occured 1 times (0.079%)
If previous instruction was djn:
DAT occured 128 times (22.980%)
SEQ occured 127 times (22.801%)
JMP occured 100 times (17.953%)
MOV occured 50 times (8.977%)
SNE occured 46 times (8.259%)
SPL occured 38 times (6.822%)
DJN occured 21 times (3.770%)
ADD occured 11 times (1.975%)
JMN occured 10 times (1.795%)
SUB occured 6 times (1.077%)
MUL occured 2 times (0.359%)
JMZ occured 2 times (0.359%)
patterns are examined up to 5 instructions deep, which means the fifth
instruction will look like:
If 4 ins ago was spl:
If 3 ins ago was spl:
If 2 ins ago was spl:
If 1 ins ago was spl:
ADD occured 1 times (50.000%)
MOV occured 1 times (50.000%)
So the fifth instruction would be an ADD or a MOV.
The modifiers and the operands are tied only to the current instruction they
are being selected for. This is done to keep the size of the tables down to a
minimum.
When creating the tables for selecting random instructions, a group of
successful warriors (I used Koenigstuhl's LP) is selected and converted to the
format that exhaust requires. I then run a program called statcode.pl, this
program counts every instruction, tallies all the start instructions it sees,
makes a note of which instructions follow which instructions, and tallies
which modifiers and operands are seen with any given instruction. The start
instructions are tallied separately to ensure that all warriors start with a
valid instruction. The program then outputs a C source file, which I compile
and link into makepop and client programs.
The source file defines several calls, get_start_instruction, get_operands,
get_imod, get_omod, and get_{second,third,fourth,fifth}_instruction.
get_start_instruction returns an instruction selected randomly from the list
of valid start instructions, weighted by the number of occurances. So in the
above example, we would expect nearly 40% of our start instructions to be SEQ,
and nearly 10% to be SPL.
The following is a source snippet from makepop.c showing how a warrior is
filled using these chains (c is a pointer to the code section of the warrior):
getstartins(&(c[0]));
op = c[0].in;
getsecondins(&(c[1]),op);
op2 = c[1].in;
getthirdins(&(c[2]),op,op2);
op3 = c[2].in;
getfourthins(&(c[3]),op,op2,op3);
op4 = c[3].in;
getfifthins(&(c[4]),op,op2,op3,op4);
for(i = 5; i < max_length; i++){
getfifthins(&c[i], c[i - 1].in, c[i - 2].in, c[i - 3].in, c[i - 4].in);
}
A second pass would then be made to mate each of the instructions with operands
and modifiers.
This method has made a significant difference in the time required to generate
a viable start population. The instruction calls are also used to suggest
mutations, reducing the chance of destructive mutations by using possibilities
that fit chains used to generate the warrior.
allowing for a small chance that we will change some part of the instruction
based on the chains that match the warriors code.
If you would like to look at the code, it is currently available for download
at www.evil-empire.info/files/ccai-1.0.tgz
Thank you to Joonas, John, Will 'Varfar' and Christian for all their ideas.
Will, thanks for the friendly competition, but I think you're behind a couple
points right now :) Joonas, thank you for repeatedly rewriting exhaust based
on my whims.
_______________________________________________________________________________
Extra Extra - Muffineer by Christian Schmidt
Here is my latest idea for a paper-stone for the LP hill. The big difference
compared with Achillu 94 or Skinner is, that the bomb run and the replication
are within the same loop. It bombs and replicate at the same time. This seems
an advantage when fighting against scanner.
;redcode-lp
;name Muffineer
;author Christian Schmidt
;strategy stone/replicator clone
;assert 1
;optimax pap
pStep1 equ 5229
pStep2 equ 3704
pAdd equ 963
pOff equ (pHit+pAdd)
zero equ qbomb
qtab3 equ qbomb
pRef mov #9, #9
pGo mov <pRef, <pBck
pPtr mov <pAdd, }pOff
add.ab {0, }0
pHit mov pBmb, @pPtr
jmn pGo, pRef
spl @pRef, <pStep2
pBck jmz @0, pStep1
pBmb dat }7506, >7279
for 85
dat 0, 0
rof
qbomb dat >qoff, >qc2
dat 0, 0
dat 0, <qb1
qtab2 dat 0, <qb2
dat 0, <qb3
for 16
dat 0, 0
rof
dat zero - 1, qa1
qtab1 dat zero - 1, qa2
for 42
dat 0, 0
rof
qc2 equ ((1 + (qtab3-qptr)*qy) % CORESIZE)
qb1 equ ((1 + (qtab2-1-qptr)*qy) % CORESIZE)
qb2 equ ((1 + (qtab2-qptr)*qy) % CORESIZE)
qb3 equ ((1 + (qtab2+1-qptr)*qy) % CORESIZE)
qa1 equ ((1 + (qtab1-1-qptr)*qy) % CORESIZE)
qa2 equ ((1 + (qtab1-qptr)*qy) % CORESIZE)
qz equ 2108
qy equ 243
qgo sne qptr + qz*qc2, qptr + qz*qc2 + qb2
seq <qtab3, qptr + qz*(qc2-1) + qb2
jmp q0, }q0
sne qptr + qz*qa2, qptr + qz*qa2 + qb2
seq <qtab1, qptr + qz*(qa2-1) + qb2
jmp q0, {q0
sne qptr + qz*qa1, qptr + qz*qa1 + qb2
seq <(qtab1-1), qptr + qz*(qa1-1) + qb2
djn.a q0, {q0
sne qptr + qz*qb3, qptr + qz*qb3 + qb3
seq <(qtab2+1), qptr + qz*(qb3-1) + (qb3-1)
jmp q0, }q1
sne qptr + qz*qb1, qptr + qz*qb1 + qb1
seq <(qtab2-1), qptr + qz*(qb1-1) + (qb1-1)
jmp q0, {q1
sne qptr + qz*qb2, qptr + qz*qb2 + qb2
seq <qtab2, qptr + qz*(qb2-1) + (qb2-1)
jmp q0
seq >qptr, qptr + qz + (qb2-1)
jmp q2, <qptr
seq qptr+(qz+1)*(qc2-1),qptr+(qz+1)*(qc2-1)+(qb2-1)
jmp q0, }q0
seq qptr+(qz+1)*(qa2-1),qptr+(qz+1)*(qa2-1)+(qb2-1)
jmp q0, {q0
seq qptr+(qz+1)*(qa1-1),qptr+(qz+1)*(qa1-1)+(qb2-1)
djn.a q0, {q0
jmz.f pGo, qptr + (qz+1)*(qb2-1) + (qb2-1)
qoff equ -87
qstep equ -7
qtime equ 14
q0 mul.b *2, qptr
q2 sne {qtab1, @qptr
q1 add.b qtab2, qptr
mov qtab3, @qptr
qptr mov qbomb, }qz
sub #qstep, qptr
djn -3, #qtime
jmp pGo
end qgo
-- Christian Schmidt
_______________________________________________________________________________
The Hint - Further Code snippets by Christian Schmidt
For warriors which contain a multi-process paper, or a vector launched imp, it
is neccessary to create a number of processes running in parallel. Beppe Bezzi
(Binary CODE) and John Metcalf (ALT 1 CODE) already discussed in their hints
possible code snippets to achieve multi-processes. In this sense I would like
to show some more snippets which sometimes can save an instruction.
The magic behind the 'new' snippets is the following single instruction:
spl 0, }0
The creates three parallel processes instead of two in the case of a spl 1.
A disadvantage is that it creates processes ahead of the rest.
N Binary CODE ALT 1 CODE ALT 2 CODE
2 spl 1, <xxx
3 spl 1, <xxx spl 2, <xxx spl 0, }0
mov -1, 0 spl 1, <yyy
4 spl 1, <xxx
spl 1, <yyy
5 spl 1, <xxx spl 2, <xxx spl 0, }0
mov -1, 0 spl 2, <yyy mov asd, 0
mov -1, 0 spl 1, <zzz (asd spl 1)
6 spl 1, <xxx spl 0, }0
mov -1, 0 spl 1, <xxx
spl 1, <yyy
7 spl 1, <xxx spl 1, <xxx spl }0, <xxx
spl 1, <yyy spl 1, }0 spl 0, <xxx
mov -1, 0 spl 1, <yyy
8 spl 1, <xxx
spl 1, <yyy
spl 1, <zzz
9 spl 1, <xxx spl 2, <xxx spl 0, }0
mov -1, 0 spl 2, <yyy mov -1, 0
mov -1, 0 spl 1, }0 spl 1, <xxx
mov -1, 0 spl 1, <zzz
10 spl 1, <xxx spl 1, <xxx spl 0, }0
mov -1, 0 spl 1, <yyy mov 1, 0
mov -1, 0 spl 2, }0 spl 1, <xxx
mov -1, 0 mov -2, 0
11 spl 1, <xxx spl 1, <xxx spl 0, }0
mov -1, 0 spl 1, }0 spl 1, <xxx
spl 1, <xxx spl 2, <yyy mov -1, 0
mov -1, 0 spl 1, <zzz
Also interesting snippets are the following two, which creates 7 and 8
parallel processes. Both are using a table which is implemented in their code.
From the dat instruction only the 1 in the a- or b-field is neccessary.
The code will be launched at pGo.
7 parallel processes:
dat 1, 0
pGo spl *-1, }0
spl 1, 0
8 parallel processes:
dat 0, 1
pGo spl @-1, }0
spl 1, 0
I think there are much more snippets to discover. Maybe we can show next time
some more.
-- Christian Schmidt
_______________________________________________________________________________
Extra Extra - Goin' Mobile by anonymous
Goin' Mobile is a "bomb dodger" that detects recent changes
to the core. When it finds a change, it assumes that the cell
was attacked by its opponent and that the cell won't be attacked
again anytime soon. That spot is considered safe by Goin' Mobile,
so it relocates there and then resumes operation.
How does it know when a cell has been altered recently?
It lays a trail of DATs as part of its normal attack plan; later,
it checks that trail for changes.
;redcode-94
;assert 1
;name Goin' Mobile
;strategy dodge stones, forward clears, and some scanners
;strategy (e.g. He Scans Alone, Blur 2)
pointer equ (top + (top-first))
top dat -100, 0
move sub.ab top, #-100
mov top, pointer
mov @top, }pointer ; copy self over footprint
jmn.f -1, >top
jmp *top, 0 ; jump to the new copy
first mov.ba move, top
wipe mov top, <move ; clear fast, scan slow
mov top, <move
jmz.b wipe, {top ; check for fresh "footprints"
jmp move, 0
end wipe
If you are fighting Goin' Mobile and you have a hard time
hitting it, then consider changing your pattern of attack
to be less predictable. For example, instead of clearing the core
in the usual order (1, 2, 3, ...), you could use MUL to calculate
the attack sequence.
;redcode-94
;assert CORESIZE==8000
;name Random Dwarf
dwarf mov what, <where
mov what, <where
where mul.ab #7961, #78
jmp dwarf, 0
what dat 1, 1
end dwarf
-- anonymous
_______________________________________________________________________________
Extra Extra - Sith by Christian Schmidt
There were introduced in the past some successful paper-stones, which are
using different 'mechanisms' to achieve nice bombing papers. Christian Schmidt's
unheard-of embedded a pure stone into the paper while David Houston's Netpaper
followed up Paul Kline' Paper Dreaming which achieves a nice bombing pattern by
co-operation of all paper copies.
A more primitive way to achieve a nice bombing pattern early in the battle looks
quite similar to a jmp/add imp launcher. It uses a spl 2 to ensure alternate
execution of the mov and add instructions.
pap1 spl @0, }pStep1
mov.i }-1, >-1
spl @0, }pStep2
mov.i }-1, >-1
spl 2, <pStep3
mov.i bOff1, {bOff2
add.ab {0, }0
pEnd djn.f -4, <pStep4
Keep in mind that we are running several parallel processes.
After constant optimization, adding a quickscanner and additional paper/imps
for a better defence it finally looks as shown below:
;redcode-94nop
;name Sith
;author Christian Schmidt
;assert CORESIZE==8000
;strategy qscan -> silk/stone & silk/imps
zero equ qbomb
qtab3 equ qbomb
iStep equ 1143
pStep equ 7342
sStep equ 5965
pStep1 equ 182
pStep2 equ 2007
pStep3 equ 7549
pStep4 equ 958
bOff1 equ 7505
bOff2 equ 4568
pAwa1 equ qbomb + 4594
pAwa2 equ qbomb + 5475
qbomb dat >qoff, >qc2
dat 0, 0
paper spl 1, <qb1
qtab2 spl 1, <qb2
spl 1, <qb3
mov.i <1, {1
spl pAwa1, pEnd+1
mov.i <1, {1
jmp pAwa2, imp+1
pap1 spl @0, }pStep1
mov.i }-1, >-1
spl @0, }pStep2
mov.i }-1, >-1
spl 2, <pStep3
mov.i bOff1, {bOff2
add.ab {0, }0
pEnd djn.f -4, <pStep4
for 4
dat 0, 0
rof
dat zero-1, qa1
qtab1 dat zero-1, qa2
pap2 spl @0, }pStep
mov.i }-1, >-1
spl #0, 0
mov pBomb, >pPtr
add.x imp, pPtr
pPtr jmp imp-iStep*8, >sStep-6
pBomb dat >1, }1
imp mov.i #sStep-1, iStep
for 34
dat 0, 0
rof
qc2 equ ((1 + (qtab3-qptr)*qy) % CORESIZE)
qb1 equ ((1 + (qtab2-1-qptr)*qy) % CORESIZE)
qb2 equ ((1 + (qtab2-qptr)*qy) % CORESIZE)
qb3 equ ((1 + (qtab2+1-qptr)*qy) % CORESIZE)
qa1 equ ((1 + (qtab1-1-qptr)*qy) % CORESIZE)
qa2 equ ((1 + (qtab1-qptr)*qy) % CORESIZE)
qz equ 2108
qy equ 243
qgo sne qptr + qz*qc2, qptr + qz*qc2 + qb2
seq <qtab3, qptr + qz*(qc2-1) + qb2
jmp q0, }q0
sne qptr + qz*qa2, qptr + qz*qa2 + qb2
seq <qtab1, qptr + qz*(qa2-1) + qb2
jmp q0, {q0
sne qptr + qz*qa1, qptr + qz*qa1 + qb2
seq <(qtab1-1), qptr + qz*(qa1-1) + qb2
djn.a q0, {q0
sne qptr + qz*qb3, qptr + qz*qb3 + qb3
seq <(qtab2+1), qptr + qz*(qb3-1) + (qb3-1)
jmp q0, }q1
sne qptr + qz*qb1, qptr + qz*qb1 + qb1
seq <(qtab2-1), qptr + qz*(qb1-1) + (qb1-1)
jmp q0, {q1
sne qptr + qz*qb2, qptr + qz*qb2 + qb2
seq <qtab2, qptr + qz*(qb2-1) + (qb2-1)
jmp q0
seq >qptr, qptr + qz + (qb2-1)
jmp q2, <qptr
seq qptr+(qz+1)*(qc2-1),qptr+(qz+1)*(qc2-1)+(qb2-1)
jmp q0, }q0
seq qptr+(qz+1)*(qa2-1),qptr+(qz+1)*(qa2-1)+(qb2-1)
jmp q0, {q0
seq qptr+(qz+1)*(qa1-1),qptr+(qz+1)*(qa1-1)+(qb2-1)
djn.a q0, {q0
jmz.f paper, qptr + (qz+1)*(qb2-1) + (qb2-1)
qoff equ -87
qstep equ -7
qtime equ 14
q0 mul.b *2, qptr
q2 sne {qtab1, @qptr
q1 add.b qtab2, qptr
mov qtab3, @qptr
qptr mov qbomb, }qz
sub #qstep, qptr
djn -3, #qtime
jmp paper
end qgo
-- Christian Schmidt
_______________________________________________________________________________
Questions? Concerns? Comments? Complaints? Mail them to people who
care. Beppe Bezzi <giuseppe.bezzi@galactica.it>, Philip Kendall
<pak21@cam.ac.uk>, Anton Marsden <anton@paradise.net.nz>, John Metcalf
<grumpy3039@hotmail.com> and Christian Schmidt <fizmo_master@yahoo.com>
|
