ARMS-D-Request@XX.LCS.MIT.EDU (Moderator) (11/11/86)
Arms-Discussion Digest Tuesday, November 11, 1986 2:11PM
Volume 7, Issue 57
Today's Topics:
Meteorite as A-explosion
First Strikes, Verifiability
Launch on warning / nuclear victory metrics
Eager retaliation ("prompt response", isn't that a home
Launch on warning
Yet more on SDI (Star Wars flawed #8-of-10)
SDI
Corrections on my factual uncertainties (2 msgs)
----------------------------------------------------------------------
From: hplabs!pyramid!utzoo!henry@ucbvax.Berkeley.EDU
Date: Mon, 10 Nov 86 19:42:14 pst
Subject: Meteorite as A-explosion
> ... Besides, the
> conversion of 1 gram of anti-matter (a cube less than 1 cm on a side)
> to energy would produce 9 x 10^20 ergs of energy, which is probably
> enough to split the earth in two...
Nothing so drastic; the erg is a pretty small unit. 1 gram of antimatter
plus 1 gram of normal matter would give about a 40 kiloton explosion, as
I recall.
Henry Spencer @ U of Toronto Zoology
{allegra,ihnp4,decvax,pyramid}!utzoo!henry
------------------------------
From: rutgers!meccts!meccsd!mvs@seismo.CSS.GOV
Date: Tue, 11 Nov 86 03:36:03 EST
Subject: Re: Arms-Discussion Digest V7 #51
Reply-To: meccsd!mvs@seismo.CSS.GOV (Michael V. Stein)
In article <8611092151.AA18053@ucbvax.Berkeley.EDU> ARMS-D@XX.LCS.MIT.EDU writes:
>
>Subject: on the perfectability of SDI
>Date: 07 Nov 86 19:45:33 EST (Fri)
>From: dm@bfly-vax.bbn.com
> 2) What is a likely Soviet reaction to our building SDI AND
> retaining our 40,000 warheads aimed at them?
The US has approximately 26,000 nuclear warheads. Of these about
13,000 could hit the Soviet Union. The Soviet arsenal is thought to
consist of somewhere between 22,000 and 33,000 warheads. (Data from
"Nuclear Battlefields").
>Put yourself in the shoes of a Russian leader. There is the USA, with
>an SDI system that is clearly only partially effective, and will
>almost undoubtedly let several hundred of your warheads through. Yet
>they also have a lot of MX and Trident missiles that really look like
>they're first strike weapons, plus those Pershing-2 missiles just 10
>minutes from your capital.
A first strike will almost certainly involve a counter-force strike.
The goal of a first strike is to eliminate the enemies ability to
fight back with his military. Trident missles are simply not accurate
enough for this sort of mission and shouldn't be included in this
list.
>...Is it cheaper
>and more reliable than the options? Considering that one of the
>options is a verifiable treaty reducing arms, I don't think so.
Great, but name just *one* time that we have had a verifiable treaty
with the Soviet Union in regards to arms control. A true verifiable
treaty must imply on-site verification, something the Soviets have
always rejected. They have explicitly rejected the concept in the
Baruch plan, in Eisenhower's Open Sky's proposal, etc.
Like just about all major US nuclear programs, SDI was started mostly
because of actual or suspected Soviet work in the field. And because
there is no true verification, this is the way the world works.
---
Michael V. Stein
Minnesota Educational Computing Corporation - Technical Services
UUCP ihnp4!meccts!mvs
------------------------------
Date: Tuesday, 11 November 1986 00:32-EST
From: Clifford Johnson <GA.CJJ at forsythe.stanford.edu>
To: LIN
Re: Launch on warning / nuclear victory metrics
REPLY TO 11/10/86 19:49 FROM LIN@XX.LCS.MIT.EDU: Re: nuclear victory metrics
I suppose I would risk everything for fear that we'd get nuked out.
Look, that is the problem with nukes in the first place. If you
believe in deterrence, it means that that you have to be willing to
promise that you will do something irrational *if* deterrence fails
and war breaks out.
But before any (other) first-use, deterrence hasn't failed? After
first-use, I agree my case becomes weaker (though I still hold to
it). But I'm adamantly opposed to operating any LOWC prior to
first-use. More precisely, I see the operation of a LOWC as a form
of first-use.
What's your position on the first-use argument?
To: LIN@XX.LCS.MIT.EDU
------------------------------
Date: 11 Nov 86 08:47:51 EST (Tuesday)
From: MJackson.Wbst@Xerox.COM
Subject: Eager retaliation ("prompt response", isn't that a home
From Herb Lin (V7 #54):
"It is precisely because. . .I judge the odds of such an occurrence to
be very low that I want to be very sure that I have solid and conclusive
evidence. If nuclear attack were a very likely thing, I might require
less evidence."
But of course the danger is that, to all appearances, at least some of
the individuals most intimately involved with strategic nuclear
command-and-control do *not* view a Soviet attack as a low-probability
event, and that they *might* require less evidence than would otherwise
be reasonable. Apropos of this whole issue, does anyone know if delayed
response scenarios are "losers" in the Pentagon's strategic wargaming?
(This would seem to bear on the observation made a while back by someone
that LUCA [Launch Under Confirmed Attack] was promptly subsumed under
simple LUA [Launch Under Attack]. This was not explained at the time; I
took the comment to mean that the use of the word "confirmed" only
served to emphasize the ambiguity likely to be present in a real-world
situation, hence to call the concept into question. How about LUAWRRSA
[Launch Under an Attack We're Really Really Sure About]?)
Mark
------------------------------
Date: Tue, 11 Nov 1986 08:49 EST
From: LIN@XX.LCS.MIT.EDU
Subject: Launch on warning
From: Clifford Johnson <GA.CJJ at forsythe.stanford.edu>
I suppose I would risk everything for fear that we'd get nuked out.
Look, that is the problem with nukes in the first place. If you
believe in deterrence, it means that that you have to be willing to
promise that you will do something irrational *if* deterrence fails
and war breaks out.
But before any (other) first-use, deterrence hasn't failed? After
first-use, I agree my case becomes weaker (though I still hold to
it). But I'm adamantly opposed to operating any LOWC prior to
first-use. More precisely, I see the operation of a LOWC as a form
of first-use.
That's why I want to maintain an OPTION for LOW, but not a POLICY.
The two ARE different; having an option means that the other guy must
worry that you might, but making it not a policy means that you can
avoid most of the risks.
I believe deterrence to have failed when missiles are on the way to
the U.S. -- and I don't want to get into the argument that sensors
could be faulty. HOWEVER I make the judgment that missiles are
coming, that is sufficient for the purposes of this argument.
I don't see the operation of LOWC at all as a form of first use. You
could argue (I would not) that it is a form of threatened first use,
but that's not the same thing.
------------------------------
Date: Monday, 3 November 1986 08:06-EST
From: Jane Hesketh <jane%aiva.edinburgh.ac.uk at Cs.Ucl.AC.UK>
To: ARMS-D
Re: Star Wars flawed #8-of-10
The Limitations of Artificial Intelligence
and the Distinction between
Special-Purpose and General-Purpose Systems
Alan Bundy
Abstract
The Battle Management System (BMS) for SDI will be a huge
computer program which can never be tested in a real battle
but only in a simulated one. Critics have argued that it
will, therefore, be inherently unreliable - reacting
unpredictably to unforeseen circumstances.
Can this limitation be overcome by the incorporation of
artificial intelligence (AI) technology, so that the BMS
will react sensibly to the unforeseen? We argue that such a
facility is so far beyond the capabilities of current AI
technology as to be unrealisable in the development period
of SDI.
A Short History of AI
The areas of AI that one might hope would provide techniques
for coping automatically with unforeseen circumstances are
the representation of knowledge and automatic reasoning with
this knowledge.
When it comes to automating the processes of reasoning,
AI has had most success in the area of deduction, i.e. the
kind of reasoning involved in proving mathematical theorems
in a formal logic. Even here success has been limited to
fairly straightforward theorems due to the problems of
controlling the search for a proof through the explosively
large space of legal steps. Forays have also been made into
the areas of: uncertain reasoning, analogical reasoning,
default reasoning, and a few other kinds of `plausible
reasoning' - but these areas are only in their infancy.
When it comes to representing knowledge, AI has had some
success in the representation of properties of and
relationships between simple physical objects. However, a
large number of tricky areas remain, e.g. the representation
of: shape, time, liquids, beliefs of others, etc, which we
are only beginning to understand. In addition, there is the
sheer scale of the problem. No existing AI system contains
more than a minute fraction of the knowledge, especially
common-sense knowledge, known to the average human.
This has limited AI systems to small, self-contained
domains in which fairly crude reasoning is adequate. Early
AI work concentrated on `toy' domains, the most famous of
which was the `blocks world' - a domain of children's bricks
in which the principal action was stacking. Until the mid-
70s AI researchers were pretty depressed about this. It was
felt that the initial predictions about AI progress had
proved to be hyperbole, that we had only scratched the
surface of the problem, and that major breakthroughs were
required. This is still true.
What changed the climate of opinion from depression to
enthusiasm was the realisation, during the 70s, that many
commercially important domains were `toy' ones in the
technical sense described above. This led to the
development of expert systems. For the most part, expert
systems use automated reasoning techniques that were already
well understood 10 to 20 years ago, and apply them to a
narrow domain of specialised expertise in areas such as
fault diagnosis. They are not capable of the kind of
common-sense reasoning about shape and time involved in
navigating a car in a busy street, nor of jumping to
plausible conclusions about the intentions of an adversary,
nor of using analogy to apply a wide range of previous
experience to a new problem. Expert systems are special-
purpose.
Is Battle Management a `Toy' Domain?
For an AI-based BMS to cope with unforeseen circumstances it
would have to be a general-purpose system. It might have to
reason about the nature, purpose and destination of
previously unknown objects. It might have to reason about
the intentions of an adversary, taking into account the
general political situation. It might have to use analogy
to cope with an unforeseen situation by adapting an existing
plan. Thus battle management is not the sort of `toy'
domain that lends itself to expert systems technology.
This is not to say that a rule-based BMS could not be
built - it could, but it would be a special purpose system
not able to cope with unforeseen circumstances. One could
regard the problem of detecting and reacting to a missile
attack as one of fault diagnosis and correction and adapt an
existing expert system shell to the task. But the resulting
BMS would be subject to precisely the same criticisms about
its reliability as a conventional BMS - in fact, more so.
The behaviour of an expert system is inherently less
predictable than that of a conventional program. Because
the rules can be combined in many different ways by the
inference mechanism according to the circumstances, the
order of rule firing may be different from any anticipated
by the programmer. This unpredictability is increased if
certainty factors are used to influence the search strategy
(as they usually are).
There is some hope that by using logic-based rules and
checking the meaning of each rule, that the program would
behave correctly whatever order its rules were fired in.
Such a program would serve as its own formal specification.
But it would be no easier to get this specification correct
than it would to correctly specify a conventional program.
In fact, it would be harder, since this specification has
the additional constraint of simultaneously serving as a
(logic) program.
Could a General-Purpose BMS Ever be Built?
The objections raised above are based on current AI
technology. What are the chances that new developments in
AI will overcome them and make it possible to develop a
general-purpose BMS?
It is an article of faith among most AI workers that it
is possible to understand general-purpose, common-sense,
human reasoning to the extent that it can be automated,
although some critics of AI (e.g. Dreyfus) argue that that
can never be done. What is universally agreed is that there
are some major theoretical breakthroughs required in the
development of knowledge representation and plausible
reasoning techniques before it can be achieved. AI
Researchers have been striving to make these breakthroughs
since the 50s. Progress has been made, but it is slow and
piecemeal. The situation is not analogous to VLSI (say),
where the basic theory is understood and hence rapid and
steady improvements in size and speed can be reliably
predicted, provided sufficient money is made available. New
theory is required and it is impossible to guarantee results
by pouring in money (*) or by any other means.
Even were the breakthroughs to occur and the AI-based
BMS to be built, how could we be sure that it would respond
`correctly' to unforeseen circumstances? Thompson has
pointed out that before we empower humans (e.g. a judge) to
take decisions which affect other people we subject them to
tests not only of technical expertise but also of basic
humanity. The latter is usually implicit and largely
assumed on the basis of a shared human experience of
upbringing and emotion. That is, we tend to assume that the
candidate shares a common morality, responsibility,
humanity, etc, unless this is undermined by his or her
actions. No such assumption would be reasonable in a
computer program which did not share human experience. How
could we test for it?
____________________
(*) Money is necessary, of course, but it is not suffi-
cient.
Conclusion
Under these circumstances it would be folly to predicate the
success of a multi-billion dollar programme on the timely
occurrence of several major breakthroughs in AI.
Information about the author
Dr Alan Bundy is a Reader in the Department of Artificial
Intelligence at the University of Edinburgh. He has been a
researcher in artificial intelligence since 1971 and has a
major international reputation in the field. He is
currently the Chair of the Board of Trustees of IJCAI Inc,
the body which organises the major AI conference. He is
Conference Chair of the next such conference, IJCAI-87. He
is also a member of the SERC Computer Science Sub-Committee,
which reviews SERC research grants in AI, and is on the
editorial board of the foremost AI Journal. He has published
many research papers and books in AI, including papers on
its nature and methodology.
------------------------------
Date: Tue, 11 Nov 1986 10:45 EST
From: LIN@XX.LCS.MIT.EDU
Subject: SDI
From: cfccs at HAWAII-EMH
The arguements against SDI seem to want the funding cut. They want a
nice slow progress that will assure no quantum jumps in technology.
One that will leave plenty of money for the projects they can't seem
to get funds for.
By what criterion would you decide what is the proper level of funding?
If you aren't
against SDI R&D (notice the D stands for developme), what are you
arguing about?
I AM arguing against SDI D, to the extent that it violates the ABM
treaty. I am arguing that SDI R should not be conducted at the level
that it currently enjoys.
If your real concern is that money will be wasted on
elaborate shows of outdated technology, then argue that!
It is not only that, though it is that. It isn't "outdated"
technology, but rather meaningless demonstrations of technological
fluff -- i.e., technology that isn't meaningful but that looks good to
TV cameras.
------------------------------
Subject: Corrections on my factual uncertainties
Date: 11 Nov 86 12:26:42 EST (Tue)
From: dm@bfly-vax.bbn.com
Hank Walker (Hank.Walker@gauss.ece.cmu.edu) has kindly corrected some
of my factual errors in my recent postings on ``waiting a while'' and
the imperfectability of SDI.
First, on waiting a while. Someone had said that taking as long as 24
hours to confirm that the Soviets had attacked would be absurd, that
it would be the equivalent of unilateral disarmament on our part,
since the Soviets would be able to use that 24 hours to completely
destroy our forces, including tracking down and eliminating all of our
nulcear submarine force. I pointed out that missing just ONE of our
Trident submarines would leave ~400 warheads for retaliation, and that
I presumed Pentagon planners had taken that thought into account and
had provided for it. Hank sent me a message correcting my numbers (in
my defense, I made it clear that I was uncertain about the numbers in
my original posting -- I'm not a professional at this, I just do it as
a hobby):
A Trident submarine carries 24 Trident I C4 missiles, each probably
carrying 8 warheads. The Trident II D5 missile under development
would carry 10 warheads over a longer range. A Poseidon submarine
carries 16 missiles. These missiles are either Poseidon missiles,
which can carry up to 14 warheads, but probably carry only 10
warheads, or C4 missiles with 8 warheads. Poseidon missiles are
gradually being replaced by C4 missiles on some fraction of the fleet.
...a Trident carries 24 missiles, currently with 8 warheads each, 10
in the future. I believe these are Mark 12A warheads with 300
kilotons, same as the Minuteman III and MX. If they aren't on the C4
missile, they will be on the D5 missile. The Poseidon missile carries
10 warheads of about 80 kilotons each. The US is allowed a total of
656 SLBMs, or 656*8 = 5248 warheads. The plan is for 20 Trident and
11 Poseidon subs, out of the total of 41 allowed subs. The D5
missiles have the added advantage that their range is sufficient to
hit targets from almost anywhere in the northern hemisphere or Indian
Ocean, so the subs will be essentially invulnerable.
I once read that the routine was that if a sub didn't hear the Moffet
Field VLF stay-alive signal for three days, it would sample the air for
radioactivity, and then launch against its target list for that
situation. This is the US's ultimate dead-man switch. Of course all
submarines have the capability of launching their missiles whenever the
crew feels like doing it.
There is of course the issue of civil defense evacuation by the
Soviets, but if nuclear winter is true, it won't matter anyway since
their first strike would probably be sufficient to cause it.
So, Tridents only carry 190-240 warheads, Poseidons about the same
number (hmm, I thought one of the arguments against the Trident was
that, under SALT-II limits, because the Trident carried more warheads
it meant we were limited to fewer platforms, meaning fewer subs for
the Soviets to hunt down. I guess I was wrong. Of course, if you
could put Poseidon missiles with their 14 warheads into a Trident
submarine, you'd get 320 warheads on a Trident...). Nevertheless, I
stand by my original assertion that missing just one submarine would
result in unacceptable (even to the inventors of the Gulag) damage to
the Soviet Union (read: the end of the Soviet Way of Life). After
all, once upon a time (mid-Sixties, before MIRV) Robert MacNamara
estimated that we'd only need a couple-hundred warheads, period.
All this goes to argue that the notion of holding back our retaliation
for a few days is a perfectly ``reasonable'' course of action.
The second correction Hank sent regarded the Pershing II missile,
which I had described as a threat to the Russian capital:
Everything I have read indicates that the Pershing II missile does not
have the range to hit Moscow from Germany. It can reach into the
western part of Russia or the Ukraine.
Oh well, so drop the Pershings as a first-strike threat. That still
leaves us with the MX missile and the Trident missile, neither of
which would I want aimed at three quarters of my nuclear forces and my
command and control centers. Particularly if the other side has an
SDI capable of mopping up what the Tridents and MXs missed.
------------------------------
Date: Tue 11 Nov 86 13:03:31-EST
From: Herb Lin <LIN@XX.LCS.MIT.EDU>
Subject: Re: Corrections on my factual uncertainties
From: <dm@bfly-vax.bbn.com>
Of course, if you could put Poseidon missiles with their 14 warheads
into a Trident submarine, you'd get 320 warheads on a Trident...).
My understanding is that you can load 14 warheads onto a Trident I or
a Trident II, but that we have chosen not to do so.
The second correction Hank sent regarded the Pershing II missile,
which I had described as a threat to the Russian capital:
Everything I have read indicates that the Pershing II missile does not
have the range to hit Moscow from Germany. It can reach into the
western part of Russia or the Ukraine.
Oh well, so drop the Pershings as a first-strike threat.
Maybe. The *Soviets* claim that the Pershing II *can* strike Moscow.
Maybe true, and maybe false, and maybe both. I have heard that one
U.S. missile carried ballast to shorten its range in order to not
count against some range limit.
Besides, there are ICBM launch control centers in the Western S.U.
------------------------------
End of Arms-Discussion Digest
*****************************