ARMS-D-Request@XX.LCS.MIT.EDU (Moderator) (09/25/86)
Arms-Discussion Digest Thursday, September 25, 1986 1:38AM
Volume 7, Issue 20
Today's Topics:
Autonomous Weapons
JANUS
----------------------------------------------------------------------
Date: Wed, 24 Sep 86 13:11:57 PDT
From: Clifford Johnson <GA.CJJ@Forsythe.Stanford.Edu>
Subject: Autonomous Weapons
Let's agree that:
AN AUTONOMOUS WEAPON IS A SET OF DEVICES PRECONFIGURED TO
CONDITIONALLY EXECUTE A BELLIGERENT ACT.
As a matter of logic, all weapons since the sword are to a greater or
lesser degree autonomous. A computer is a machine, and machines have
been around for awhiles. The following was written well before the
advent of computers:
"Through a blur of dust and fumes there appear, quite suddenly, great
black and yellow masses of smoke which seem to be tearing up the
surface of the earth and disintegrating the works of man with an
almost visible hatred. These form the chief parts of the picture,
but somewhere in the middle distance one can see a few
irrelevant-looking human figures, and soon there are fewer. It is
hard to believe that these are the protagonists in the battle. One
would rather choose those huge substantive oily black masses which
are so much more conspicuous and suppose that the men are in reality
their servants, and playing an inglorious, subordinate, and fatal
part in the combat. It is possible, after all, that this view is
correct." (J.B.S. Haldane, in Deadalus, or Science and the Future,
1924.)
Because of the real sense in which all modern weapons are deemed
autonomous, it is fruitful to turn to a definition of the (degree of)
autonomy of an autonomous weapon:
A WEAPONS SYSTEM IS AUTONOMOUS TO THE DEGREE THAT IT IS PRECONFIGURED
TO CONDITIONALLY EXECUTE BILLIGERENT ACT(S).
Thus, the autonomy of a weapon is constructed from the weapon's
"condition space," which is the set of conditions governing the
execution of the belligerent act(s), and from the "outcome space,"
which is the set of the weapon's possible act(s), only some of which
may be belligerent, and into which the weapon is configured to map
the conditions. (Note that much of this model applies to autonomy in
general.) The analysis of a weapon's autonomy can accordingly
proceed with the formal description of: (a) the condition space; (b)
the outcome space; and (c) the mapping function. Each is is
important to the overall assessment of autonomy, and proneness to
error would be an important consideration within each topic.
For example, missiles guided by embedded processors are more
autonomous than bombs guided by gravity due to the greater complexity
of the condition and outcomes spaces, and of the mapping from the
first to the latter. However, due to the greater variability in the
outcome set (explode/don't vs. explode at A/B), a mine may qualify as
more autonomous than a guided missile, despite the relative
complexity of the missile.
The development of a typology of condition spaces, outcome spaces,
and mapping processes would generate a typology of autonomies. For
example, autonomy would be essentially "digital," "continuous," or
"hybrid," depending on whether the condition space had these
characteristics. Other important classifications of the condition
space might include whether the conditions were "pre-adjustable"
(e.g. by alert-levels) or "target-acquisition (goal) oriented". The
outcome space might could be "disparate" or "compact," depending on
the variance between possible outcomes, and the mapping processes
might be "utility-driven," or "goal-seeking."
Prodigious advances in microscopic digital technology have led to the
development of weapons systems an order of magnitude more autonomous.
It would be surprising if this growth in automation were not
associated with fresh cases of illegality under international law.
These cases sometimes fall within the scope of extant legal
categorizations, but may add fresh emphases due to the qualitatively
new encroachment upon decisionmaking domains hitherto reserved for
human discretion, and the enlargement of the overall domain of
decisionmaking capabilities.
The law of nations is manifestly not silent upon the legitimacy of
autonomous weapons. Thus, a particular type of autonomous weapon may
be illegal persuant to consequent probabilities of conflict
initiation, disproportionate response, innocent fatalities, treaty
violation, or other infringement of international law. Indeed,
autonomous weapons, notably mines and other booby-trap devices, have
received special legal attention. The ability of computers to mimic
more complex decisionmaking situations than arise with simple
booby-trap designs has led to an urgent need to extend the
established categories of prohibition. In particular, it should be
pronounced by authoritative legislative and judicial councils that,
with respect to autonomous weapons:
THE NATURE OF A BELLIGERENT ACT AND OF THE CONDITIONS AND CONTEXT OF
ITS EXECUTION MAY INHERENTLY REQUIRE RESPONSIBLE DISCRETION AT THE
TIME THE CONDITIONS GOVERNING EXECUTION ARE EVALUATED. THIS
RESPONSIBILITY REQUIREMENT IS NOT NECESSARILY MET BY MERE PROVISION
OF REAL-TIME OVERRIDE CAPABILITY.
There already is a body of international and military law which
supports this contention. Delegation, authorization, and competence
doctrines may be especially pertinent. Consider, for example,
the case presented by the Strategic Defense Initiative, which
proposes weapons preconfigured to open fire based upon digitized
sensor inputs...
To: ARMS-D@XX.LCS.MIT.EDU
------------------------------
Date: Wed, 24 Sep 86 15:20:58 pdt
From: eugene@AMES-NAS.ARPA (Eugene Miya)
Subject: Re: JANUS
Background, I first heard of the JANUS some years back in Time in an
article about computer gaming. There was one very poor photo (color)
of a terminal. The accompanying article said that (publicity value)
one officer so flubbed his game that he took the largest nuclear
weapon in his stockpile, dropped it where he thought the enemy was,
and succeeded in destroying both the enemy and his forces.
JANUS is a teaching tool for the use of tactical nuclear weapons and
other battle field management. JANUS is located at the Conflicit
Simulation Center at LLNL. Everything I mention is unclassified, and
I was informed that an article appeared during mid-August in the SJ
Merc News about Janus. (have not seen it.)
The system runs on a VAX-11/780 (VMS v4.2) with 8 tektronix color
graphics workstations with graphics tablets (4 button), or microVAXen
workstations. 4 workstations constitute the Blue Force and 4
constitute the red Force. Workstations have telephones which can be
used to simulate radio communications [I found this a bit hokie].
Playing the game on a 1200 baud, or a 9600 baud or lately, a 19.2K
baud modem is significant: the faster the modem, the fast a side can
respond.
The system is largely menu driven. My demonstration had 5 companies
(A-E) and the menu had chemical and nuclear options (both of which
were disabled since I think this was an unclassified demo). A typical
"game" lasts only a day, but the planning for that one day takes a
minimum of three days.
The system is designed for middle-range commanders from Captains to
Colonels. These people typically have no computer experience and work
with people familar with the system who "program movement orders,
etc." The important part is the communications and cooperation a
commander has with his subordinates. Playing this game: 1) it's like
batch rather than interactive programing. You don't want to
interactively hack a battle, you can't keep up. You want small
changes like directing artillery, not big movements. Good battles
have well set-up plans to flexibility handle a wide variety of
situations. Again, it's designed as a teaching tool rather than a
full blown policy (tactics) simulator. The computer handles low-level
details, but humans are typically always playing against humans. (My
demo was against the machine with a preprogramming Red force, and
limited human intervention, see below, about helicopter movement).
You clearly don't want a colonel telling every single man below him
what to do.
My guide was a helicopter gunship pilot by training.
> Yes, I would like a description of what it can do.
There are a variety of sources of intelligence (some visual, others sensors)
information, there is combat engineering (anti-tank trenches mine fields),
status information (casualities, reserve units), graphics zoom.
> What does it simulate?
The simulations starts with a topographic map and an area. There are
icons representing tanks, APCs, helicopters, forward observers, etc.
My area was at Ft. Irwin in the Mohave desert about an 8 x 8 mile area
on the Fort.
It simulates "the physical characteristics of a system." This includes
the effective ranges of cannon, missiles, and so forth. It includes
limited weather characteristics. Example: we fired artillery to lay down
smoke to cover the retreat movement of an armored unit. The cloud
slowly dispersed during the course of the game. It's not weather
ala the global circulation programs. Don't ask for rain or fog in
great detail. Running surfaces are simulated to varying degrees.
The same thing can be said about night combat.
Once engagement starts, it simulates rounds flying throughout a
battlefeld.
It's not programmed for amphibitious or naval operations, only land and
limited air operations (helicopters and A-10s).
> How are real data put into it?
Real data is keyed in INGRES. Limited interaction takes place using
key board and tablet: Movement information, fire control.
The system is not "object-oriented." Enabling units (platoons)
was cumbersome. There are some real user-interface problems:
menus were strictly top-down and you have to pop up levels to get out.
Programming evasion must be interesting since general movement orders
can be programmed in.
There is no use of `AI technique.' There is limited smarts: if a
movement order is near enough to a road, the tank will drive on
the road.
> What level of simulation does it do?
It's a tactical simulation of units from the platoon up to brigade or
battalion level. You can see individual tanks if you want to.
It's an intermediate physical/command simulation.
> What kind of weapons hardware does it handle?
We clicked buttons on both the Blue and Red forces, and I asked questions.
There were table entries for things like M60A3 tanks, M113 APCs, BMPs,
T-64s, M-1s, T-72s, Cobras, Dragon launchers, TOWs, Hellfire, howitzers,
and the unspecified nuclear and chemical options (I tried a chemical option,
no-go, and they told me that it was nothing specific: simply a non-descript
non-persistent, not GB).
From a given weapons platform, you could ask for line of sight information,
weapons ranging info, you could set movement orders.
Example: I had two helicopters which I activated as part of the demo.
I moved them to maximum range of the armored column where they started
to pick the spearhead apart (Intelligence could not tell me where the tail
was). We then moved to the Red force terminal, and changed the movement
orders of the Red team Anti-Aircraft unit to move wide right
where they eventually shot the helicopters down.
> How realistic is it?
Physical:
I've been to Fort Irwin where our topographical simulation took place,
so I've seen the tanks and have driven the one of the roads the Blue
force was defending (This is where the Goldstone Tracking Station is
located). The objective to for a Red Force armored column to take a
Blue Force defended pass and road intersection. The forces were
equally matched, but Blue was more thinly spread since it was playing
defensively. The simulator would have a hard time deailing with
forest information. It's not for street fighting in cities either.
It's realism has some severe limtations. Helicopter air support only
flies at one altitude. This is an obvious simplification, more
levels, would add more computational requirement.
The game does not stop. Taking status information can cost you.
Status information is displayed directly over the playing field, so
if something significant gets covered you lose.
Non-physical: There are big questions about how lower-level firing
decisions are made (the probablistic ones: like hold fire and attack
the tail of the column).
Part of the intersting realism is the confusion during a battle. All
decisions are made on incomplete information. I could occasionally
look at the Red force screen only because one partition was down. The
moment intelligence information about Red forces appeared on a screen,
all hell broke lose. On this stand point, it was real.
Realism is also very dependent on the programming. A force will
suffer 90% casualities without blinking an eye. Retreats must be
specifically programmed, there is no rule otherwise. Suicide missions
ala Apple II games are very real.
Conclusion of my game: after WE shot down Blue's two helicopters, I
was given control of Blue forces and I was overrun. Red force (what
was left) took over my intersection. I still had scattered intact
Blue forces miles behind Red-lines. We are trying to get the CSC
people to give an open technical talk locally.
The invitation to visit the JANUS facility and see a demo came to
me under the aegis of ACM SIGGRAPH but not directly to my employer
who supports my SIGGRAPH activity. I wish to thank SIGGRAPH for
the opportunity, and the people at LLNL who took the time and effort
to show me their facility.
I'll send a description on the use of computers in the design of
nuclear weapons in November.
From the Rock of Ages Home for Retired Hackers:
--eugene miya
President
Bay Area Chapter, Assoc. for Comp. Mach.
Special Interest Group on Computer Graphics (SIGGRAPH)
NASA Ames Research Center
eugene@ames-aurora.ARPA
"You trust the `reply' command with all those different mailers out there?"
{hplabs,hao,dual,ihnp4,decwrl,allegra,tektronix,menlo70}!ames!aurora!eugene
------------------------------
End of Arms-Discussion Digest
*****************************