ARMS-D-Request@XX.LCS.MIT.EDU (Moderator) (11/08/86)
Arms-Discussion Digest Saturday, November 8, 1986 1:58PM Volume 7, Issue 52 Today's Topics: Military Institutions and AI Yet more on SDI (Star Wars flawed #3-of-10) Mid-course Interceptions 24 hour waiting period? ---------------------------------------------------------------------- Date: Sat, 8 Nov 1986 13:26 EST From: LIN@XX.LCS.MIT.EDU Subject: Military Institutions and AI From: toma at Sun.COM (Tom Athanasiou) Does anyone know of institutional forces within the military that predispose positive receptions for technologies that don't really work. Yes. Promotions are often based on the budget one controls or on the visibility of the program. The peacetime military has little incentive to develop weapons that actually work. A program goes on, and develops bugs. What happens then? What program manager is going to say "This won't work" when his neck is on the line? What is the incentive for him to do so? If early in the program, he says "It's too early to tell what will happen." If late in the program, he says "Look at all we have spent on the program -- all of that will be wasted if we stop." There's been a lot of talk about SDI, but I'm interested in AI per se. The level of hype in the commerical AI world has dropped a lot faster than in the military AI world. Why? In the commercial world, there is a bottom line -- whether something does its job. In the peacetime military, there is no comparable bottom line. Does anyone know of anyone that would be helpful to talk to on this issue? Of anything that would be good to read? Fallows' "National Defense" is a good place to start, though you should not take his word as gospel. Steubing's "The Defense Game" is also pretty good. For people to talk to, let me know what you are interested in. ------------------------------ Date: Monday, 3 November 1986 08:04-EST From: Jane Hesketh <jane%aiva.edinburgh.ac.uk at Cs.Ucl.AC.UK> To: ARMS-D Re: Yet more on SDI (Star Wars flawed #3-of-10) Fully automatic decision-making systems Henry Thompson It has become clear in recent years that computer technology is as crucial to strategic weapons systems as physics is. Leaving aside the terrifying possibility of one of the nuclear powers moving to a fully automatic launch on warning or launch under attack policy, the clearest example of this is the suggested role of computer systems in the SDI Battle Management System (BMS). In the Fletcher report to the US Department of Defense, it was stated that response times would be so rapid as to preclude significant human participation: "The battle management system must provide for a high degree of automation to support the accomplish- ment of the weapons release function." This is clearly in fact an under-statement - for effective boost-phase response the decision time is certainly less than two minutes, probably less than one, even without fast-burn boosters - and it is clear that the authors of the Eastport report expect at least the boost-phase BMS to operate without any human participation. What this means then is that at least during crisis periods when the system was fully enabled, the decision that a hostile attack was underway and that an active response should be initiated, together with the orchestration of at least the early stages of that response, would be inaccessible to human intervention. This moves us immediately into the domain of Artificial Intelligence, but at a level so far beyond current experience as to be difficult to imagine. In an effort to understand just what the deployment and empowerment of a fully automatic SDI BMS would mean, it is worth looking for a moment at where we are now in this area. The answer is emphatically nowhere. No existing or announced AI system has been empowered to act independent of human participation. All the existing expert systems of which so much is said function in an advisory, not an executive, capacity. The most complex and sophisticated fully automatic systems which involve computers in use today are barely worthy of the description `decision making', for example traffic signal controllers, cash dispensers and autopilots. All the systems which have been suggested as possible models for the SDI BMS, including the control systems for the Apollo moon flights, the phone network and the various space probes are not in this list, because none of them exhibit(ed) any significant autonomous decision- making. What then is an automatic decision-making system? What indeed is a decision? Consider a room thermostat, responsible for controlling the operation of a heating system in response to variations in temperature. We can diagram its essential properties as follows: ___________________________ heating ___________________________ temp off on ___________________________ high nothing turn off ___________________________ low turn on nothing ___________________________ | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | That is, a thermostat must discriminate among four possible sorts of situation, and act accordingly to turn the heating on or off or leave it alone. These notions of discrimination and action are constitutive of a decision- making system. What then is meant by an automatic decision-making system? Simply one that, like a thermostat, involves no human participation, either to guide the discrimination or to approve the action. Thermostats, foghorns, burglar alarm and traffic signal controllers are all examples. There are of course degrees of automation. Systems like those for zero-visibility aircraft landing and load-balancing in the national power grids are subject to human supervision, but of a fairly minimal and post-hoc nature. There are also degrees of involvement of computers, but above a certain level of complexity electro-mechanical ingenuity fails and only computational approaches are possible. Even thermostats have microprocessors in them by now, in the more sophisticated cases. There is a crucial step in the deployment of automatic decision-making systems, which has not so far received the attention it deserves. This is the step of empowerment, that is, the point at which control passes to the system. Before we empower human beings such as teachers, pilots, policemen or judges to make decisions we (that is, society acting through government) typically subject them to a more or less strict regime of training and evaluation. We will clearly soon reach, if we have not already, the stage with artifacts, that is with automatic decision-making systems, where explicit controls on empowerment will be required. What tests might one wish to perform on an artifactual candidate for empowerment? How would one go about determining its fitness for its appointed task? Many of today's criticisms of the SDI programme can be seen as reasons why there is no possibility of a sufficiently convincing evaluation of a candidate SDI BMS to make empowerment a responsible action. The question of empowerment is a new one - it has not arisen before because the necessary combination of sophisticated technology and human impact has only recently emerged. We cannot allow SDI to be exempted from a general requirement for a sensible empowerment process for automatic decision-making systems, yet it is clear it could not `pass' such a review. Consequences of failure - active versus passive defense The empowerment issue for SDI gains tremendous significance when the nature of the proposed system is examined carefully. In a crucial way the briefly mooted name of Peace Shield was misleading. SDI is not a passive defense, like a shield or an umbrella. It is an active defense, like a fly-swatter or an anti-aircraft gun. To be effective it must be wielded, it cannot just sit there. This means it is liable to two sorts of failures. Like a passive defense, it might, as it were, leak. The overwhelming evidence, tacitly acknowledged even by the SDIO, is that any SDI we can hope to build will fail this way. But unlike a passive defense, it might also be wielded in error. All our experience to date of supervised computer decision-making systems suggests this will in fact happen. The literature is full of examples of computer errors, either in specification or implementation, provoking false alarms, with disaster averted only by human intervention. But ex hypothesi in the case of SDI no such intervention would be possible. If at some point, perhaps in the midst of an international crisis, the system was fully empowered, then if owing to some flaw in specification or implementation it incorrectly determined that a missile attack was underway and began its response, there would be no possibility of human intervention. A vast array of weapons would be unleashed, some harmlessly, at non-existent missiles, perhaps a forest fire or a meteor shower, but some at Soviet satellites. What chances then that anyone would survive the doubtlessly automatic response to that very real assault, to examine the program and correct the flaw? In a fully automatic system there is no escaping the fundamental law of signal detection - there is a trade-off between type 1 and type 2 errors, between misses and false alarms. You cannot eliminate one without guaranteeing that the other will occur. A safe SDI would have to be so constrained as to be useless - it would almost surely not work when it was needed. A useful SDI would have to be sufficiently unconstrained as to be unsafe - it would be too likely to `work' when it was not needed. Glossary BMD Ballistic Missile Defense BMS Battle Management System C3I Command, Control, Communications and Intelligence LOW Launch on Warning LUA Launch under Attack MOU Memorandum of Understanding (between US and UK on SDI) SDI Strategic Defense Initiative SDIO Strategic Defense Initiative Organisation (in the US Department of Defense) SDIPO Strategic Defense Initiative Programme Office (in the UK Ministry of Defense) Information about the author Henry Thompson is a lecturer in the Department of Artificial Intelligence and the Centre for Cognitive Science at the University of Edinburgh. Before coming to Britain in 1980 he was a member of the Natural Language group at the Xerox Palo Alto Research Center. He has been doing research in the areas of knowledge representation and computational linguistics for more than 10 years, and has a long-standing interest in the philosophical foundations of Artificial Intelligence. He is currently co-director of the Edinburgh part of an Alvey Large Scale Demonstrator project whose goal is an interactive, incremental speech input system. ------------------------------ Date: Sat, 8 Nov 1986 13:47 EST From: LIN@XX.LCS.MIT.EDU Subject: Mid-course Interceptions From: crummer at aerospace.ARPA >Another concept, discussed in candidate architectures, is one in which >you do very effective mid-course discrimination without doing >boost-phase intercept, perhaps using interactive discrimination with >neutral particle beams. With the miserable signal/noise ratio that a sensor would have looking at a cool target against a star field, how would these particle beams be pointed? I'm sure the beam can't be fired and steered at the rates necessary to "paint the sky" and find the objects even if interactive discrimination would work. Herb, do you have any more information on the mid-course fantasy? The beams would probably be pointed on the basis of some space-based radar or IR sensor looking at the threat cloud; information on objects would be passed to NPBs for discrimination. Painting the sky is absurd, and SDIO knows that. ------------------------------ Date: Sat, 8 Nov 1986 13:57 EST From: LIN@XX.LCS.MIT.EDU Subject: 24 hour waiting period? From: Paul F. Dietz <DIETZ%slb-test.csnet at RELAY.CS.NET> Someone has advanced the position that the US should wait 24 hours after a nuclear attack before retaliating. A response was that this would eliminate US bomber forces. The proponent claimed US bombers could stay up this long, pointing to the fact that NEACP (sp?) can stay up three days. Well, I looked that up; the command plane can stay up for *at most* three days, given sufficient in-air refueling. The three day time limit comes from the engines running out of lubricant. After a large nuclear strike there would likely be little in-air refueling capability left. I spoke of 24 hours, not of 72. Besides, the 3 day limit is in no way fundamental. If we wanted to equip bombers with oil replenishment systems, we could. I did not say that the entire bomber force could survive, just that some of it could. This business about whether or not all the bombers could survive and whether or not the scenario proposed in which the Soviets attack us and we are left with only submarine bases is starting to miss the point of waiting. There is certainly a down side to waiting. I don't think that is it is as great as critics have said (for example, some bombers not in the air WOULD survive at auxiliary airfields etc), but I willingly concede a down side in terms of reduced effectiveness and increased costs. The issue is NOT whether or not there is a down side. The issue is that by waiting (and incurring the costs of waiting), you may *increase* the probability that you won't react mistakenly to an attack report. My original query was of the form "Describe a scenario in which waiting the U.S response would be significantly impeded." I haven't received one yet, but for the sake of argument let's assume there is one. The relevant question is "What is the likelihood of that scenario?" and how does it compare to the likelihood of receiving a false attack report? Both are dangers that we must consider. ------------------------------ End of Arms-Discussion Digest *****************************