Jim, Thanks for your comments. However, I do have to disagree with your cheerful assessment of the chances for REALISTIC testing of any SDI system or major subsystem. The problem is not that there is no testing that could be done, but that any substantial change in the system's environment is likely to provoke a new set of unexpected behaviors. Certainly, an SDI that failed to shoot down a meteor swarm could be judged faulty. (And certainly any experienced software engineer would predict that it WOULD fail to shoot down its first meteor swarm.) But what reason is there to believe that an SDI that had shot down every meteor swarm since its deployment would act in the intended manner when faced with a full-scale nuclear attack, which would certainly be accompanied by both attacks on the SDI system itself and by extensive counter-measures? It is in the nature of counter-measures that you cannot be sure in advance that you have full knowledge of the counter-measures that your opponent will throw at you--especially in a first engagement. Thus, there is in principle no way to adequately test your system's response to counter-measures. It is in the nature of distributed and real-time systems that the most catastrophic failures tend to come in periods of heaviest load. Thus the results small-scale testing can't be extrapolated with confidence. It is in the nature of destructive testing (which a full-scale nuclear attack on an SDI certainly would be) that you can't test the thing that you will ultimately rely on. However, it had better be something so assuredly similar to the ultimate system that you can be confident in the extrapolation. Finally, one brief non-software point: I read a column the other day (Tom Wicker?) that pointed out that every technology that would be effective against ballistic missiles would be far more effective, far sooner, as an anti-satellite system, since the behavior of satellites is more predictable, and the attacker can pick his moment. Even without SDI, the US is far more dependent on satellites for both defense and civilian uses than the USSR. And any space-based SDI would make a very tempting ASAT target. So developing the technology needed for the system INCREASES the risk of relying on the system. Jim H.
Cc: nrl-css!horning@decwrl.ARPA The comments that you reported on testing SDI, proposing that we test it by shooting at periodic meteor swarms make me wonder how many of the people in our profession have trouble discriminating between real problems and arcade games. Shooting at an easily predictable non-evasive meteor has about the same relation to the real job of SDI as shooting at a target in a booth at a county fair has to shooting at woods in heavy brush from a moving airplane. If I had a computer program that had been tested by controlling a weapon at a county fair, I would not have much confidence in its performance the first time that the B.C. Government tried to use it in its periodic wolf kills from light planes. In fact, I hope someone sells that idea to them. Dave
From: Herb Lin <LIN@MC.LCS.MIT.EDU> [Date: Sat, 4 Jan 86 18:52:04 EST] [Herb's message is embedded. PGN] From: Jim McGrath <MCGRATH at OZ.AI.MIT.EDU> The model to think of is a sophisticated computer game. The human operator(s) would take care of truly strange cases (rising moons, flocks of interplanetary geese).. But the major problem is not the things that the computer isn't sure about, but rather the things that it is sure about that are not true. How would the human ever know to intervene? I thought a bit about that, and have a suitable elaboration. Basically, you require a "two key" system, with the computer holding one key and a human operator/monitor another. This is primarily for the "go/no go" decision. After an attack is acknowledged, you concede the possibility of overkilling by the computer (taking out third party satellites and the like) in return for the more immediate response to attack provided by the computer. This takes care of the computer going off half cocked. If you are worried about the computer missing an actual attack, you can now set the sensitivity low, trusting to the human monitor to not activate when appropriate. Actually, this is too simple. What you really want is to have the hardware/software under a set of human operators, perhaps partitioned to provide zone coverage. The humans act as before, mainly as checkpoints for activation decisions, overseeing strategy, sending expert information to the computers as the situation unfolds so that the software does not have to be a tactical genius. Now a set of human supervisors sit on top of the operators. They have another "key," and so can break ties on activation decisions (or even override lower level decisions). Their other missions are to advise operators on developing strategy, keep the command authorities informed, and to act as "free safety." That is, they will have the authority to override operator commands so that targets that find seams in the zones (or similarly defy the operator/computer teams) will be targetted for attack. Normally they will access information at a much higher level than an operator (the former will have to deal with thousands of targets - the latter tens of low hundreds). Other concepts can be advanced: advance/retard the ease of a go/no go decision according to alert status and the like. The main point is that a man in the loop is a big win, since you get a proprogrammed general purpose computer which can take care of those "higher level" decisions. Response time is not a concern - seconds are not vital if you have 20 minutes. Only for boost phase interception do you run into difficulties. Jim
From: Herb Lin <LIN@MC.LCS.MIT.EDU> ... "go/no-go" You mean fire/don't fire? I mean weapon activation. Firing decisions for specific targets will be made by computer, but the weapons themselves will be inert until activated. After an attack is acknowledged, you concede the possibility of overkilling by the computer (taking out third party satellites and the like) in return for the more immediate response to attack provided by the computer. So your solution is that you kill everything, and don't do discrimination? No. I meant exactly what I said. You concede that you might make a mistake in firing (which was your original objection). You do not aim for making a mistake. I explicitly said in the same message that one of the jobs of human operators is to assist in real time parameter adjustment so that the computer controlled weapons would be able to discriminate better. As I said earlier, boost phase poses a particular problem. The only thing I can see to do now is to trust in AI to give you a good initial screen, and to argument this with a human authorized to override the problem in a few seconds. This could work well for limited periods of time (such as alerts), but I have problems with it for extended periods. Jim
From: Jim McGrath <MCGRATH at OZ.AI.MIT.EDU> This [full-scale system testing — HL] seems to be a common problem with any modern weapon system (or even not so modern - it took WWI for the Germans to realize that the lessons of the 1880's concerning rapid infantry fire (and thus the rise of infantry over calvary) did not take artillery development adequately into account). And there have been disasters. Only here, the disaster is bigger. What if, after suitable advance notice, the SDI system was fully activated and targeted against one of our periodic meteor swarms? While not perfect targets, they would be quite challenging (especially with respect to numbers!), except for boost phase, and CHEAP. If the system was regenerative (i.e. you only expended energy and the like), then the total cost would be very low. Interesting example, but problematic. No kill assessment for one, under some circumstances. Entirely different signatures for another. Meteors are just a casual example. My point is that the costs of partial (but system wide) testing does not have to lie with the targets (which many people seem to assume) as much as with weapons discharge - which may be quite manageable. But if the tests are to be realistic, then the right targets are essential, especially since a counter-measure is to try to fool with
From: Jim McGrath <MCGRATH at OZ.AI.MIT.EDU> remember that the major cost of the target simulation is in the boost phase. Once the targets are in sub-orbit, it makes no difference whether they were fired independently by hundreds of expensive boosters or were accelerated from orbital velocity, after having been place there originally through more economical means. Terminal phase tests are especially easy to do this way. Only boost phase is intrinsically expensive. I agree with your technical point. But successful boost phase is what SDI is all about. The technology for dealing with mid-course and terminal is ALREADY here. You need boost phase so that you can thin out the midcourse and terminal.
From: Herb Lin <LIN@MC.LCS.MIT.EDU> From: Jim McGrath <MCGRATH at OZ.AI.MIT.EDU> What if, after suitable advance notice, the SDI system was fully activated and targeted against one of our periodic meteor swarms? Interesting example, but problematic. No kill assessment for one, under some circumstances. Entirely different signatures for another. It would test some aspects of the system on a system wide level (such as detection and tracking), and would even provide good kill estimates in some cases (KE weapons and small targets). But as I said: Meteors are just a casual example. My point is that the costs of partial (but system wide) testing does not have to lie with the targets (which many people seem to assume) as much as with weapons discharge - which may be quite manageable. But if the tests are to be realistic, then the right targets are essential, especially since a counter-measure is to try to fool with the targets that the defense sees. True, but remember that the major cost of the target simulation is in the boost phase. Once the targets are in sub-orbit, it makes no difference whether they were fired independently by hundreds of expensive boosters or were accelerated from orbital velocity, after having been place there originally through more economical means. Terminal phase tests are especially easy to do this way. Only boost phase is intrinsically expensive. (That's two messages where I've come up with approaches to problems that work on all phases except boost phase. Although initially attractive, perhaps concentrating more on mid-course and terminal defense will ultimately prove more beneficial.) Jim
> From: Herb Lin <LIN@MC.LCS.MIT.EDU> <> From: horning at decwrl.DEC.COM (Jim Horning) <> More generally, I am interested in reactions to Lipton's proposal that <> SDI reliability would be improved by having hundreds or thousands of <> "independent" orbiting "battle groups,".. > That is absurd on the face of it. To prevent propagation of failures, > systems must be truly independent. To see the nonsense involved, > assume layer #1 can kill 90% of the incoming threat, and layer #2 is > sized to handle a maximum threat that is 10% of the originally > launched threat. If layer 1 fails catastrophically, you're screwed in > layer #2. Even if Layers 1 and 2 don't talk to each other, they're > not truly independent. True but his solution WOULD reduce the probability of the propagation of "hard" errors (i.e. corrupting electronic communications), and the whole independence approach should lead to increased redundancy so as to deal with "soft" propagation of errors such as you cite. Remember, you do not need to PREVENT the propagation of errors, just reduce the probability enough so that your overall system reliability is suitably enhanced. I think the approach has merit, particularly over a monolithic system, and should not be shot down out of hand. This is the fundamental point of disagreement. If SDI is just another defensive system, then all that you say is right. But it isn't. I will stop beating the perfect system horse when the SDI supporters acknowledge that large-scale population defense can never be made certifiably reliable.
What do you mean by "certifiably reliable?" While politicians may talk about 100% reliability, we are all scientists and engineers here - we know that nothing, including such things as the sun rising, is 100% reliable. You must really mean X% reliable, where X is a high number (perhaps high enough so as to reduce to a very low probability the chance of a single warhead getting through). In that case, independent battlestations, and other measures, might give you the number you need. I submit that it is too early to dismiss these approaches out of hand, since you are really talking about a quantative difference and we do not have good numbers yet. Anyway, I am arguing for a highly reliable, but by no means perfect, system. My X would probably be lower than yours. I really do think that there is a difference between a few million dead (horrible, on the scale of WW II) and hundreds of millions dead (utterly unprecedented). And while I am certain that we all, including the public, would like as high an X as possible, they would agree that losing a city or two and some missile bases/airfields would be a lot better than losing everything. Besides, complaints that politicians are lying do not sit well with me. Of course they are lying. WE WANT THEM TO LIE. Politicians who tell the truth get kicked out of office. Our entire posture of extended deterrence is a joke, since we do not have the capability to creditably back it up. But you try to get someone elected promising to reinstate the draft, raise the defense budget further, or pull back our troops and cut Europe/Japan loose. We have to make do with what we have. Jim [Do we have to have it just because we can make do with it? PGN]
From: Jim McGrath <J.JPM at Epic> What do you mean by "certifiably reliable?" A system whose performance is known in advance to be adequate to the task. I don't care if the number for reliability isn't 100%, just high enough so that no one dies. We all, including the public, would like as high an X as possible, they would agree that losing a city or two and some missile bases/airfields would be a lot better than losing everything. But that is not the goal of the SDI. Besides, complaints that politicians are lying do not sit well with me. Of course they are lying. WE WANT THEM TO LIE. [...] We have to make do with what we have. So you condone lying to the public as a tool of public policy? How would you like to acknowledge that publically in a letter to the NY Times? Don't forget to add that you support SDI, and that truth doesn't matter when you try to justify a weapon system — never mind what it actually does. We can say that we will spend millions of dollars on AIDS research since that will save lives, and spend the money instead on nerve gas, which will also help to eliminate AIDS (by killing homosexual soldiers). Sorry; I believe that elected leaders have a responsibility to tell the truth to the public, and to educate them away from fairy tales. I would rather see precious defense dollars go to create good anti-tank weapons; that would have some chance of improving extended deterrence.
From: Herb Lin <LIN@MC.LCS.MIT.EDU> We all, including the public, would like as high an X as possible, they would agree that losing a city or two and some missile bases/airfields would be a lot better than losing everything. But that is not the goal of the SDI. Which does not mean it should not be supported for that reason. Most government programs have consequences (sometimes good, sometimes bad) never conceived of in their initial purpose. That does not mean you ignore them when evaluating the program. I simply do not follow your logic at all. Do you want to score points against Reagan and Company? Or do you want to discuss strategic defense, and SDI as it is developing? I'm not interested in defending Reagan, just developing defense and seeing that it is done the best way possible. Jim
From: Herb Lin <LIN@MC.LCS.MIT.EDU> Besides, complaints that politicians are lying do not sit well with me. Of course they are lying. WE WANT THEM TO LIE. Politicians who tell the truth get kicked out of office.... So you condone lying to the public as a tool of public policy? [...] You are arguing from emotion (almost hysterically), not reason, which I do not expect of you. I stated a fact: public officials must lie on many (not all) issues in order to retain office. (I could have said "evade," or just "keep quiet about" if the word "lie" hits you so hard - I see no functional distinction.) This is one thing that everyone, no matter what their policy perspective, agrees on (this comes from several graduate seminars, and personal experience). I did not say that I liked that state of affairs much. But I do not find it reasonable to blame the politicians. Rather, the fault lies with the voters. Unlike many of my friends in the social sciences, I do not concentrates on the "oughts" of the world. I focus on the empirical evidence. Perhaps it is the scientist in me. So when I observe a political system that punishes frank and honest talk about some issues (usually those, like nuclear war and taxes, that are too horrible to contemplate), I acknowledge this as a fact, and do not waste time decrying it. My decrying it is not (to the first approximation) going to change human nature. Thus my comment "we have to make do with what we have." Sorry; I believe that elected leaders have a responsibility to tell the truth to the public, and to educate them away from fairy tales. I would rather see precious defense dollars go to create good anti-tank weapons; that would have some chance of improving extended deterrence. Come on now. Leaders can only lead where people are, ultimately, willing to go. Just look at the nuclear freeze movement. This is the level at which the public thinks of nuclear war when it is forced to think. Finally, your last sentence shows that you missed my entire point. Congress (i.e. the people) will not budget for the necessary increases in conventional weapons (let alone the Europeans). Ultimately it does not matter what you or I like, it is what the people will accept. And if they act "irrationally," then I feel we cannot just sit back and demand that out "leaders" make them change their minds, or that the people change their stripes. Instead we should focus on the possible - which is, afterall, what politics is all about. Jim
Not speaking legally, but just morally, I think any professional who relies on a single source of advice when more are available is derelict. In any life-critical situation, information should always be tested or cross-checked whenever possible. Would you exonerate a pharmacist (or a physician, for that matter) who relied completely on a particular reference book which had a critical error in it? Or on the tape-recorded lecture of some (human, less-than-perfect) instructor? Or, in the case under consideration, on some expert computer system? I certainly wouldn't. The pharmacist does not shoulder all the moral burden in this hypothetical case. Some of it belongs to the information source (publisher, human instructor, or expert computer system). Just not all of it. As the other commenters on the subject have noted, the user must be carefully trained against unthinking dependency. That's the single most important factor. This is another instance of the urgent necessity to debunk the popular myth of computer infallibility. People are all too eager to stop thinking and let others (parents, teachers, priests, politicians, computers) make all their decisions. And the others usually love the role, too. In the case of computer systems, we their builders are in this seductive position and we must remember the inherent perils. --Rodney Hoffman
In Volume 1, Issue 34, firstname.lastname@example.org talked about faking login procedures and non-system software masquerading as system software. Well, an interesting thing just happened to me along those lines. We've got several modems used for both dial-in and dial-out. To place an outgoing call, you disable logins on a line (a program, acucntrl does this for you) and use kermit to talk to the modem. I had just done this when somebody called in. The modem, of course, didn't know logins were disabled and answered the phone. At this point, nothing much prevented me from pretending to be a login process and prompting the unsuspecting user for his login name and password, then faking a burst of noise and breaking the connection so s/he wouldn't suspect anything out of the ordinary. Our PBX allows you to forward all your calls by dialing "12 NNN". Until canceled, all incoming calls get routed to xNNN. I could have had the modem forward all its calls to my office phone. Plug an Apple-II w/modem into my phone line, and I'm all set to steal passwords for a night. Fix it up in the morning, and nobody would suspect anything worse than a modem temporarily going bonkers, which happens often enough. Roy Smith <allegra!phri!roy> System Administrator, Public Health Research Institute 455 First Avenue, New York, NY 10016
Please report problems with the web pages to the maintainer