Issue 32 June 3, 1996
______________________________________________________________________________
Core Warrior is a weekly newsletter promoting the game of corewar. Emphasis
is placed on the most active hills--currently the '94 draft hill and the
beginner hill. Coverage will follow where ever the action is. If you have
no clue what I'm talking about then check out these five-star internet locals
for more information:
FAQs are available by anonymous FTP from rtfm.mit.edu as
pub/usenet/news.answers/games/corewar-faq.Z
FTP site is: ftp.csua.berkeley.edu /pub/corewar
Web pages are at:
http://www.stormking.com/~koth ;Stormking
http://www.ecst.csuchico.edu/~pizza/koth ;Pizza
http://pauillac.inria.fr/~doligez/corewar/ ;Planar
Newbies should check the stormking page for the FAQ, language specification,
guides, and tutorials. Post questions to rec.games.corewar. All new players
are infinitely welcome!
If ftp.csua.berkeley.edu is unreachable, you can download pMARS at:
Terry's web page--http://www.infi.net/~wtnewton/corewar/
Planar fpt site--ftp://ftp.inria.fr/INRIA/Projects/para/doligez/cw/pmars
______________________________________________________________________________
Greetings.
Not much action on either hill this week so there is little to report, but
big changes are coming to a pizza hill near you. Thos has proposed two
upgrades to the hill script: 1) eliminate self fights and 2) include the
'test' option. While the merit of self fights may still be discussed, the
addition of the 'test' option will be welcome by everyone. Not only will it
increase hill throughput by reducing full 25 challenge test warriors, but
this feature should eliminate some of the artificial aging. Other much
wanted options mentioned in the newsgroup include:
-- consolidated e-mail with all warrior results
-- updates on your warriors' score against all current warriors
-- ability to change the ;redcode line to control e-mail from the hill
-- ability to modify the ;strategy lines
These are not demands, merely wishful suggestions to two individuals--Thos
and Tuc--who donate some of their time and effort to making core war a much
more enjoyable game.
Having trouble picking that magic constant? Want to know how to most
efficiently cover core? Read Stefan Strack's short tutorial on MOPT, the
multiple constant optimizer. Also in this issue, Paul Kline discusses how
to deal with all those nasty imp/ stones that have invaded the hill lately.
Look for it in this week's hint.
--M R Bremer
______________________________________________________________________________
Current Status of the Internet Pizza Server ICWS '94 Draft Hill:
Hill Specs:
coresize: 8000
max. processes: 8000
duration: after 80,000 cycles, a tie is declared.
max. entry length: 100
minimum distance: 100
rounds fought: 200
instruction set: ICWS '94 Draft
# %W/ %L/ %T Name Author Score Age
1 42/ 37/ 21 Twister Beppe Bezzi 147 31
2 41/ 38/ 22 Blur Anton Marsden 144 323
3 37/ 31/ 32 Osseil John K W 144 76
4 35/ 26/ 40 the historian bjoern guenzel 144 50
5 42/ 41/ 17 Yogi P.Kline 143 2
6 34/ 26/ 40 Scotch Broth 1.0 Robert Macrae 143 40
7 38/ 34/ 28 stoninc Maurizio 142 170
8 43/ 47/ 10 d-clear II bjoern guenzel 139 8
9 39/ 39/ 22 Chameleon M R Bremer 139 1387
10 38/ 38/ 24 T.N.T. pro Maurizio Vittuari 139 295
11 37/ 37/ 25 Thermite II Robert Macrae 137 1916
12 27/ 17/ 56 Hazy Shade II John K W 137 875
13 40/ 43/ 17 Scan Man David van Dam 137 336
14 40/ 45/ 15 HopScotch 1.6 Robert Macrae 136 44
15 35/ 35/ 30 Bomber Boy David van Dam 134 205
16 28/ 22/ 49 Rosebud Beppe 134 250
17 39/ 45/ 16 Stepping Stone Kurt Franke 134 729
18 40/ 48/ 12 test P P.Kline 133 1
19 35/ 37/ 28 Grilled Octopus v0.5 David Boeren 133 1132
20 29/ 26/ 45 blue candle bjoern guenzel 132 164
21 38/ 45/ 16 Violent Micro v0.4 basehead 132 181
22 22/ 15/ 63 ompega Steven Morrell 130 177
23 23/ 18/ 59 EIF John K W 127 124
24 22/ 22/ 56 Wicked John K W 122 12
25 30/ 52/ 18 Test Steven Morrell 109 3
Weekly age: 42 ( 127 last week, 88 the week before )
New warriors: 7 Turnover/age rate 17%
Average age: 341 ( 311 last week, 435 the week before )
Average score: 136 ( 140 last week, 141 the week before )
The top 25 warriors are represented by 13 authors: Wilkinson with 4, guenzel
and Macrae with 3, and Bezzi, Morrell, Kline, Vittuari, and van Dam with 2.
Only 42 successful challenges!
d-clear reigned early in the week, until Bezzi submitted his newest version
of Twister, which utilizes a qscan before the Tornado type bomber. Kline
shortens Yogi Bear to Yogi, changes some plogic, and voila, a top five
warrior. EIF, Rosebud, blue candle, and Violent Micro have all dropped in
the standings while veterans Thermite II and Chameleon manage to gain a few
spots. Blur and Osseil have remained relatively stable, neither one dropping
out of the top ten. One time king Scan Man has been everywhere from number
2 to number 21. Hopefully there will be some more action next week or when
the new hill scripts get finished and authors take their vengence on specific
warriors.
______________________________________________________________________________
94 - What's New
# %W/ %L/ %T Name Author Score Age
21 28/ 29/ 43 Scotch Broth 1.0 Robert Macrae 128 1
1 42/ 38/ 20 Twister Beppe Bezzi 146 1
23 22/ 19/ 59 Wicked John K W 125 1
2 45/ 44/ 11 d-clear II bjoern guenzel 145 1
25 31/ 51/ 18 Test Steven Morrell 111 1
3 42/ 42/ 17 Yogi P.Kline 142 1
18 40/ 48/ 12 test P P.Kline 133 1
Twister has been near the top spot ever since it was submitted. d-clear
started off strong, but is now in the 8th slot. Yogi is still young and most
likely has a lot more testing to go before Paul is satisfied with him.
______________________________________________________________________________
94 - What's No More
# %W/ %L/ %T Name Author Score Age
26 1/ 0/ 3 007 David van Dam 5 4
26 0/ 0/ 3 Test Paper Robert Macrae 4 4
26 2/ 1/ 2 test rtiqe1 Beppe 6 57
26 1/ 1/ 1 Yogi Bear P.Kline 5 35
26 2/ 1/ 2 d-clear bjoern guenzel 7 29
26 1/ 0/ 3 Lithium And Water John K W 5 24
26 0/ 47/ 53 Unknown Anonymous 53 2
Most of these were killed by their authors.
______________________________________________________________________________
94 - What's Old
# %W/ %L/ %T Name Author Score Age
11 37/ 37/ 25 Thermite II Robert Macrae 137 1916
9 39/ 39/ 22 Chameleon M R Bremer 139 1387
19 35/ 37/ 28 Grilled Octopus v0.5 David Boeren 133 1132
12 27/ 17/ 56 Hazy Shade II John K W 137 875
17 39/ 45/ 16 Stepping Stone Kurt Franke 134 729
13 40/ 43/ 17 Scan Man David van Dam 137 336
2 41/ 38/ 22 Blur Anton Marsden 144 323
Scan Man and Blur are beginning to show some age, but have a long way to go
before they become eligible for the Hall of Fame.
______________________________________________________________________________
HALL OF FAME
* means the warrior is still active.
Pos Name Author Age Strategy
1 Impfinity v4g1 Planar 1993 Stone/ imp
2 Thermite II Robert Macrae 1916 * Qscan -> bomber
3 Jack in the box Beppe Bezzi 1620 P-warrior
4 Tornado 3.0 Beppe Bezzi 1567 Bomber
5 Torch t18 P.Kline 1539 Bomber
6 Frontwards v2 Steven Morrell 1420 One shot scanner
7 Chameleon Myer R Bremer 1387 * P-warrior
8 Evol Cap 6.6 John Wilkinson 1299 Imp / stone
9 quiz Schitzo 1262 Scanner/ bomber
10 T.N.T. Maurizio Vittuari 1204 Bomber
11 Grilled Octopus v0.5 David Boeren 1132 * P-warrior
12 Iron Gate 1.5 Wayne Sheppard 926 CMP scanner
13 Agony II Stefan Strack 912 CMP scanner
14 Barrage Anton Marsden 876 Qscan -> replicator
17 Hazy Shade II John Wilkinson 875 * P-warrior
15 Blue Funk Steven Morrell 869 Stone/ imp
16 Flurry Anton Marsden 835 Qscan -> pwarrior
18 Thermite 1.0 Robert Macrae 802 Qscan -> bomber
19 Blue Funk 3 Steven Morrell 766 Stone/ imp
20 Night Train Karl Lewin 755 Replicator
21 Mirage 1.5 Anton Marsden 736 Scanner/ bomber
23 Stepping Stone Kurt Franke 729 * Qscan -> Vampire
22 Blizzard Anton Marsden 713 Qscan -> replicator
24 HeremScimitar A.Ivner,P.Kline 666 Bomber
25 La Bomba Beppe Bezzi 650 Qscan -> replicator
Thermite II continues its trek towards the top of the Hall of Fame standings
and the 2000 mark. Hazy Shade II advances two places and Stepping Stone
moves up one spot. The next warrior that could enter the hall is van Dam's
Scan Man, but it is still a long ways away at an age of 323.
______________________________________________________________________________
Current Status of the Internet Pizza Server Beginner's Hill:
Hill Specs:
coresize: 8000
max. processes: 8000
duration: after 80,000 cycles, a tie is declared.
max. entry length: 100
minimum distance: 100
maximum age: At age 100, warriors are retired.
rounds fought: 200
instruction set: ICWS '94 Draft
# %W/ %L/ %T Name Author Score Age
1 61/ 28/ 11 Violent Micro v0.4b basehead 194 51
2 58/ 30/ 12 Sapper Andrew Fabbro 185 25
3 51/ 31/ 18 Mace Edgar 170 19
4 49/ 31/ 21 scan test bjoern guenzel 167 79
5 47/ 31/ 22 Toxin IVa Edgar 162 94
6 49/ 36/ 15 radar.b John K. Lewis 162 48
7 46/ 32/ 22 Trebuchet Andrew Fabbro 160 32
8 43/ 26/ 31 Hunter V 0.8 O.Fechner 159 21
9 45/ 31/ 24 Toxin IVb Edgar 158 81
10 50/ 41/ 9 Drill 1.2 Edgar 158 23
11 50/ 43/ 8 Drill 1.1 Edgar 157 96
12 47/ 40/ 12 Rusty.2 John Lewis 154 4
13 44/ 34/ 22 Extremely Prejudiced Scott Manley 153 26
14 40/ 33/ 27 Fork v0.1-13p (i) Christoph C. Birk 147 6
15 41/ 43/ 16 Bloodhound Mk III Andrew Fabbro 139 89
16 39/ 44/ 17 testvss2 Iain Hogg 133 16
17 38/ 44/ 18 Yet Another Try 1a Justin Kao 133 2
18 41/ 50/ 9 Micro Centurion (mod) Edgar 132 53
19 37/ 45/ 18 testvss Iain Hogg 128 17
20 25/ 40/ 35 Hunter V 0.5 beta O.Fechner 111 28
21 24/ 39/ 36 Mutant Strain John K. Lewis 109 5
22 30/ 59/ 10 gundam John K. Lewis 101 3
23 26/ 62/ 12 Roundup Jeff Hafner 90 1
24 20/ 55/ 25 bomberQ2 harleyQ2 84 29
25 23/ 66/ 11 Test1 Oscar Strawczynski 80 9
Only 23 successful challenges! Of course it is summer vacation for most
colleges. John K. Lewis has increased his presence on the hill with some new
warriors. O.Fechner seems to have brought Hunter out of beta and is enjoying
some success in the top ten.
______________________________________________________________________________
The Hint
by Paul Kline
Beppe and others seem to be taking the 94 Hill over with those strong, fast
stone-imp warriors. So it's time for an antidote :-)
Now, there are many ways to kill stones including paper, dodgers, faster
stones, and sometimes big slow imps. One principle that works well is
numbers. A single stand-still program is at a disadvantage against an
opponent running at multiple locations. Kill one part and the others just
run faster. Even a pair of bombers running at half-speed will prove very
tough on the stone.
I was working on this concept for a couple of days with some success and just
about to submit something to the Hill for a trial when lo-and-behold Bjoern
Guenzel published the perfect component for this purpose -> d-clear.
My concept (not a new one) was to start a pair of core-clears running at a
separation of 4000. The stone would probably cripple one of them but the
other might last long enough to get him. And even with rough implementation
it worked pretty well. After the stone is wiped, any imps advance to the
forward-clear pointer and are killed as well.
I just have to say something about Bjoern's innovative use of the djn.f to
drive his forward clear. Many of us had tried the forward decrement in a
program or two, but it has mostly been abandoned for that purpose.
Unfortunately when the decrement hits it's own pointer it stops - the pointer
is both incremented and decremented in the same instruction so there is no
more progress, making it less destructive than a backward decrement. Bjoern
on the other hand saw this as a feature and used it to drive a forward core
clear at 100% of c, alternately bombing and decrementing core multiple times
until the whole thing has been dat-wiped. If the clear-pointer is hit by dat
it skips the clear code as usual. But if the decrement hits the pointer it
leaves it at zero until the dat is written there - wonderful!
Wanting to see how d-clear worked against my own pair of programs, I ran a
trial. Yikes! d-clear won about 60% of the rounds, what magic is he
using??? A close inspection shows that d-clear starts his bombing run
halfway around core from himself, meaning that the first of my components he
kills is the one nearest to killing him. Then a race to catch the other
part, but since he is running at 100% of c and mine only at 66% he stands an
excellent chance of killing that one too.
Replacing my two components with d-clear -like code gave enormous improvement
and good results against the stone-imps. Here are some examples:
Rosebud 101/37
EIF 128/31
d-clear 191/18
blue candle 91/65
impfinity 117/35 (almost 60%!)
Ompega 58/50
the historian 77/91 (my best score against him)
On the other hand:
Twister 57/124
Blur 25/164
TestBB 71/116
Thermite 61/113
Hazy Shade 47/125
So you can see that it is vulnerable to papers and any kind of stun-bomb or
continuous scanning program, which after all are designed to kill
multi-location programs. A good p-component?
So here is the code for test P, renamed 'P-clear' for publication:
;redcode-94
;name P-clear
;kill P-clear
;author P.Kline
;assert CORESIZE == 8000
;strategy pair of forward dat-wipers based on Guenzel's d-clear
;strategy outnumber stones and forward-wipe imps
paGate equ Pair-7
slPair mov #-3950,paGate+1300-7 ; initialize wipe pointers
mov #-3950,paGate+5302-7
spl 1 ; use multiple processes
spl 1
mov -1,0
mov <paBoot1,{paBoot1 ; boot 'em up
mov <paBoot2,{paBoot2
paBoot1 spl Pair+1300,Pair+7
paBoot2 jmp Pair+5302,Pair+7
Pair spl #0 ,<paGate-2667
mov paBomb ,{paGate ; slow backward clear for redundancy
mov paBomb ,>paGate ; fast forward clear
djn.f -1 ,>paGate
dat 0,0
paBomb dat 0,15
end slPair
Paul Kline
pk6811s@acad.drake.edu
______________________________________________________________________________
Extra Extra
by Stefan Strack
A brief mopt tutorial
~~~~~~~~~~~~~~~~~~~~~
Mopt, short for _M_ultiple constant _opt_imizer, is a program that simulates
a core warrior in order to find an optimal combination of bombing or
scanning constants. First of all, what is an optimal constant? An optimal
constant is a bomb/scan increment that covers core evenly, such that gaps
between adjacent bombed or scanned locations shrink at similar rates. For
example, a bomb increment of 4 is fairly bad, because it bisects core
unevenly, leaving one small and one huge gap. There are a number of
programs, e.g. Jay Han's CoreStep and Nandor Sieben's Optima, that find
_single_ optimal constants by analytical and iterative methods. These
programs work well for simple warriors like A.K. Dewdney's Dwarf. However,
most modern warriors are more complicated, dropping several bombs or
scanning several locations in each loop. Consider for example the scan loop
of a prototypical, SPL-carpet bombing CMP-scanner:
scan sub.f incr,comp
comp cmp.i FIRST,FIRST-CDIST ;larger number is A
slt.a #len,comp ;compare to A-number
djn.f scan,<FIRST-DJNOFF ;decrement A- and B-number
....
incr dat.f <-IVAL,<-IVAL
the line at 'comp' scans two locations, CDIST apart, and is incremented by
IVAL every time trough the loop. Obviously, the effectiveness of the scan
depends on chosing an optimal combination of CDIST and IVAL; this is where
Mopt comes into play. You can think of this scan pattern as consisting of
two parallel, single increment scans. The first starts scanning at location
FIRST and increments in IVAL steps; the second starts scanning at location
FIRST-CDIST and also increments in IVAL steps. As a shorthand, we can use
the notation (FIRST,IVAL),(FIRST-CDIST,IVAL) to describe this scan pattern.
Since core is cyclical, the value of FIRST is unimportant, and only the
absolute value of CDIST matters. The pattern therefore simplifies to
(0,IVAL),(CDIST,IVAL); CDIST is the "offset" of the second pattern. With
this, we're ready to start up Mopt, first, to give us a score for a single
combination of (increment,offset) values (0,42),(12,42):
Coresize: 8000
Increment value or generator #1: 42 (the offset of pattern 1 is implicitely 0)
Offset value or generator #2: 12
Increment value or generator #2: 42
Offset value or generator #3: <Enter>
Mopt stops prompting for input of (offset,increment) pairs when you input an
empty line and produces:
42 12 42 -> 35.187500
The higher the value, the better the (offset,increment) combination. This is
not terribly useful, however; you want Mopt to _find_ an optimal
(offset,increment) combination. To do this, you substitute a "value" with a
"generator", which outputs a range of values. In our example, let's try
different values for CDIST, the scan offset, in the range of 10 to 15
(shorter would be inefficient, longer would take too long to SPL-carpet).
In our notation: (0,42),([10..15],42). The generator to accomplish this is:
a=10,a<=15,a=a+1
This looks remarkably like the for loop syntax of the C-language; "a=10"
initializes variable a to 10, "a<=15" is the condition that has to be
satisfied for the loop to continue, and "a=a+1" increments the loop variable
and generates the values for optimization. Mopt uses the pMARS expression
evaluator, so you can use all its logical and arithmetic operators, as well
as assigment and 26 single letter variables. Variables are "global", i.e.
they retain their values between generators. We're now ready to start up
Mopt again:
Coresize: 8000
Increment value or generator #1: 42
Offset value or generator #2: a=10,a<=15,a=a+1
Increment value or generator #2: 42
Offset value or generator #3: <Enter>
The best:
42 12 42 -> 35.187500
42 14 42 -> 34.975875
42 11 42 -> 34.431000
42 13 42 -> 34.361125
42 15 42 -> 34.007750
42 10 42 -> 33.761125
So, a scan offset of 12 is best, followed by 14, 11, a.s.o.. Now that we
have the basics, we can let Mopt optimize _both_ scan increment and offset.
We'll pick a mod-2 pattern in the range 22 through 98:
b=22,b<99,b=b+2
Note that this generator will also produce mod-4, mod-6, .. numbers, which
we have to weed out later.
Coresize: 8000
Increment value or generator #1: b=22,b<99,b=b+2
Offset value or generator #2: a=10,a<=15,a=a+1
Increment value or generator #2: b
Offset value or generator #3: <Enter>
The best:
28 14 28 -> 44.714625
26 14 26 -> 44.443625
34 14 34 -> 42.757750
26 13 26 -> 42.690000
98 14 98 -> 42.347875
28 15 28 -> 42.242250
78 14 78 -> 42.231125
...
(Side note: If you're now wondering why Rave uses a 42,12 scan instead of
the more optimal 28,14 pattern: 42,12 turns out to be a better imp-killer).
Another example:
spl #A,#-B
add -1,1
mov <2,<2
jmp -2
The bomb pattern is (0,A),(0,-B) and we're trying to find optimal mod-5
values for A and B:
Coresize: 8000
Increment value or generator #1: a=5,a<200,b=b+5
Offset value or generator #2: 0
Increment value or generator #2: b=-5,b>-200,b=b-5
Offset value or generator #3: <Enter>
The best:
155 0 -155 -> 83.736000
155 0 -185 -> 79.708125
185 0 -155 -> 79.704375
155 0 -195 -> 79.425250
195 0 -155 -> 79.420875
155 0 -165 -> 78.837750
165 0 -155 -> 78.834625
...
Now that you understand how to translate your warrior's bombing/scanning
loop into the (offset,increment) notation, Mopt should be a snap to use.
Mopt is part of the ptools10.zip archive, and comes as portable C-source and
DOS executable. To compile it yourself, you also need eval.c and the header
files of the pMARS source distribution. The URL is:
http://www.stormking.com/~koth/pmars.html
Plans for the next version of Mopt include the ability to specify a "maximum
target size". In the current version, this is hardwired to 200, the
MAXLENGTH of the 94x hills.
-Stefan Strack (stst@idnsun.gpct.vanderbilt.edu)
______________________________________________________________________________
Questions? Concerns? Comments? Complaints? Mail them to people who care.
authors: Beppe Bezzi <bezzi@nemo.it> or Myer Bremer <bremermr@ecn.purdue.edu>
|
