Andy Freeman criticizes the following by Michael L. Scott, "Computers have no such abilities. They can only deal with situations they were programmed in advance to expect." He writes, "Dr. Scott obviously doesn't write very interesting programs. :-) Operating systems, compilers, editors, mailers, etc. all receive input that their designers/authors didn't know about exactly. " Scott's statement is not refuted by Freeman's. Scott said that the computer had to have been programmed, in advance, to deal with a situation. Freeman said that sometimes the programmer did not expect what happened. Scott made a statement about the computer. Freeman's statement was about the programmer. Except for the anthropomorphic terms in which it is couched, Scott's statement is obviously correct. It appears to me that Freeman considers a program interesting only if we don't know what the program is supposed to do or what it does. My engineering education taught me that the first job of an engineer is to find out what problem he is supposed to solve. Then he must design a system whose limits are well understood. In Freeman's terminology, it is the job of the software engineer to rid the world of interesting programs. Reliable compilers, editors, etc., (of which there are few) are all designed on the basis of a definition of the class of inputs that they are to process. We cannot identify the actual indvidual inputs, but we must be able to define the class of possible inputs if we are to talk about trustworthiness or reliability. In fact, to talk about reliability we need to know, not just the set of possible inputs, but the statistical distribution of those inputs. Dave Parnas
From: Andy Freeman <ANDY at Sushi.Stanford.EDU> Operating systems, compilers, editors, mailers, etc. all receive input that their designers/authors didn't know about exactly. When was the last time you used a mailer, operating system, compiler, etc.. that you trusted to work *exactly* as documented on all kinds of input? (If you have, pls share it with the rest of us!) It can be argued that SDI isn't understood well enough for humans to make the correct decisions (assuming super-speed people), let alone for them to be programmed. That's a different argument, and Dr. Scott is (presumably) unqualified to give an expert opinion. His expertise does apply to the "can SDI decision be programmed correctly?" question, which he spends just one paragraph on. You are essentially assuming away the essence of the problem by asserting that the specs for the programs involved are not part of the programming problem. You can certainly SAY that, but that's too narrow a definition in my view.
Herb Lin writes: When was the last time you used a mailer, operating system, compiler, etc.. that you trusted to work *exactly* as documented on all kinds of input? (If you have, pls share it with the rest of us!) The programs I use profit me, that is, their benefits to me exceed their costs. The latter includes their failures (as well as mine). A similar metric applies to weapons in general, including SDI. (Machine guns jam too, but I'd rather have one than a sword in most battle conditions. The latter are, for the most obsolete, but there aren't perfect defenses against them.) Lin continued with: You are essentially assuming away the essence of the problem by asserting that the specs for the programs involved are not part of the programming problem. You can certainly SAY that, but that's too narrow a definition in my view. Sorry, I was unclear. Specification and implementation are related, but they aren't the same. There are specs that can't be implemented acceptably (as opposed to perfectly). Some specs can't be implemented acceptably in some technologies, but can in others. (This can be context dependent.) Dr. Scott's expertise applies to the question of whether a given spec can be programmed acceptably, not whether there is an spec that can be implemented acceptably. Much of the spec, including the interesting parts of the definition of "acceptable", is outside CS, and (presumably) Dr. Scott's expertise. Another danger (apart from simplification to incorrectness) of expert opinion articles is unwarranted claims of expertise. Dr. Scott (presumably) has no expertise in directed energy weapons yet he claims that they can be used against cities and missiles in silos. Both proponents and opponents of SDI usually agree that it doesn't deal with cruise missiles. If you can kill missiles in silos and attack cities, cruise missiles are easy. -andy
This is in response to John Allred's comments about the loss of both the Thresher and the Scorpion (RISKS-3.82). Date: Mon, 20 Oct 86 13:31:40 EDT From: John Allred <email@example.com> Subject: Loss of the USS Thresher Thresher, according to the information I received while serving on submarines, was lost due to a catastrophic failure of a main sea water valve and/or pipe, causing the flooding of a major compartment. The cause of the sinking was reported by the mother ship during the boat's sea trials. Just to confirm what John stated, fracture of a hull-penetration fitting, at the weld between the flange and the pipe, quickly flooded the engineering spaces. The sinking had nothing to do with the reactor. Scorpion, on the other hand, had no observer present. No reason of loss has been given to the public. Scorpion was in very high speed transit, westbound in one of the submarine transit lanes, when she struck a previously uncharted undersea mountain. The speed of the collision was "in excess of forty miles per hour" (probably closer to sixty). It was the very high speed that had rendered her (acoustically) blind; unable to see the obstacle in her path. True, no observer was present, but a lot of people did get to hear the result. The "days spent searching for the lost sub" were just to avoid revealing how accurate our tracking capabilities were. All the Navy brass knew within the hour, exactly what had happened and exactly where.
Disclaimer: a few months ago, my knuckles were rapped when I incorrectly cited a study on airline safety. Please be warned that I know absolutely nothing about nuclear submarines and am using the ongoing discussion about automatic controls for nuclear reactors (on submarines) only as a starting place for a wider discussion. From the discussion on Risks it seems that, while automatic controls may do a satisfactory job of running the reactor in normal circumstances, people will still be needed to run the reactor when the automatic controls malfunction. Adding automatic controls adds weight (and probably noise), making the ship less effective. Adding automatic controls to a nuclear submarine's reactor frees personnel for other tasks. But, there isn't much else for them to do (they can hardly chip rust on the deck), so they'll get bored and lose their "combat readiness." Relying on totally manual control keeps the crew alert and aware of the action of the reactor. It also keeps them busy. In other words — and I think this is directly relevant to Risks — there are times when external factors make it unwise to automate a task, even when it can easily be done. Martin
While looking something up in Martin Shooman's book on software engineering yesterday, I came across the following footnote (p.495): Alan Kaplan, the editor of Modern Data magazine, posed the question, "Is the ABM system capable of being practically implemented or is it beyond our current state-of-the-art ?" The replies to this question were printed in the January and April 1970 issues of the magazine. John S. Foster, director of the Office of Defense Research and Engineering, led the proponents, and Daniel D. McCracken, chairman of Computer Professionals against ABM, led the opposition. It's startling that the very question that so interests us today was put 15 or so years ago; to make it the exact question, all you have to do is change the 3 letters of the acronym. And this was 3 (?) generations ago in computer hardware terms (LSI, VLSI, VHSIC ?) and some indeterminate time in terms of software engineering (I can't think of anything so clear-cut as circuit size to mark progress in software). International politics, however, seems not to have changed much at all. I'll try to track down those articles (Modern Data no longer exists having become Mini-Micro Systems in 1976), but in the meantime can anyone shed light on this debate from the dim past ? (BTW, Shooman comments "Technical and political considerations were finally separated, and diplomatic success caused an abrupt termination of the project." p. 498)
I agree with Anon that overall system effectiveness is non-linear: >If 1000 missiles strains the system to the point that it can only >stop 800, why would anyone think it could stop more when the number of >missiles and decoys is doubled, straining the system's ability to >identify, track, and destroy missiles at least twice as much? The more reasonable (and conservative) assumption is that the SDI system would stop ZERO missles when faced with, say, 2000 targets. Case in point is revision n of the US Navy Aegis system — seems that being designed to track a maximum of (17) targets, when there are (18) targets the computer software crashed. Any engineered artifact has design limits. When stressed beyond those limits, it fails. We understand this for civil engineering artifacts, such as bridges. Clearly this is not well understood for software engineering artifacts.
After reading the recent ARMS-D on the Stealth subject, particularly the interesting message from Bryan Fugate where he says that "stealth fighters and bombers have already gone into production", and in light of some of the recent aircraft collisions, I couldn't help but wonder if anyone has adequately considered the air traffic control consequences of not being able to get a radar fix on a large, rapidly moving aircraft in a high density air traffic area? For that matter, what about ground-radar-assisted-landing in poor visibility at a military base? Sometimes you want an aircraft to present a GOOD radar target. As I was writing this I thought of the answer, I guess. The stealth aircraft would have to have a strong beacon turned on in these circumstances. I guess it's easy to recreate a good target this way. All I can say is that the beacon had better be working in the circumstances I described.
We visited New Zealand a few years ago and went to the major skiing area on the North Island (the name escapes me). It is built on the slopes of an active volcano. There were prominent warnings for skiers of what to do in case of an eruption alarm. (Head for a nearby ridge. Don't try to outrun the likely mud/ash slide coming down the hill.) How do they get the alarm? There is an instrument hut at the lip of the crater connected to park headquarters by a cable. The instruments measure some parameter(s) or other. (heat, acceleration, pressure, ?) When something crosses a threshold, the warning alarms on the ski slopes are set off automatically. In fact, someone admitted, what would probably happen is that the explosion would destroy the hut and cut the cable. Loss of signal is probably as good a diagnostic as anything else. I can imagine a display on the DC-10 instrument panel inscribed with the outline of the aircraft. Little red lights come on when you lose continuity on a wire to an engine, aileron, etc. - like what happens when you leave your door open on a Honda Civic. What you do with this data is another matter.
Please report problems with the web pages to the maintainer