I recently had a terrible experience (involving loss of several hours of work time and much swearing) with a print spooler program for the Macintosh. As with many such programs, you put it on the system by running an installer program. The installer performs all sorts of modifications to the system software, and basically you just gotta cross your fingers and hope the software works okay. This one didn't. As soon as I attempted to print something, the system no longer saw my printer and the printer would go into an alarm condition every time I tried to print something. Calling the software company was no help, I probably spent 1/2 hour listening to an NPR station in Louisiana hoping they would pick up the phone. I tried de-installing the software, by putting the old system up from backups and tracking down all the files the installer had inserted into my system. This was a lot of work, and had no effect on the problem. This is an important point I'd like to make: any installer program should also have the capacity to de-install, so that the user has at least one back-tracking path if something goes wrong. A de-install feature probably wouldn't have helped me out, though. The problem turned out to be that the program accidentally changed one of my system defaults. Without asking, it activated Applelink (the networking system) which normally uses the line printer port. That caused my default printer port to switch to the modem port. Mark Thorson firstname.lastname@example.org
Recent discussion of problems with election equipment and software has not included any mention of the efforts of the Federal Election Commission and National Association of State Election Directors (NASED) to promote the development of standards and certification testing of election systems. NASED just recently recognized the first national Independent Testing Authority (ITA) to conduct testing of election equipment and software. Election Technology Laboratories (ETL), a joint operation of Stevens Institute of Technology and U.S. Testing Company, is now ready to conduct hardware, electrical, environmental, package, software, and human factors testing of election systems. We are very interested in learning about election system breakdowns. Since election officials do not typically advertise the breakdown of election systems, it would be good to hear from anyone in the RISKS FORUM about their experiences.
There is a great body of literature about the way people make decisions and the systematic biases in their thinking. Just to cite one example, people tend to be risk-taking when a question is framed in terms of possible gain, but risk-aversive when the question is framed in terms of possible losses. Perhaps the best single reference on this is the book "Judgement Under Uncertainty: Heuristics and Biases", edited by Kahneman, Slovic, and Tversky. (My copy is almost 10 years old, but it was published by Cambridge University Press, ISBN 0-521-28414-7). Kahneman and Tversky in particular have spent a long time analyzing the systematic biases in peoples' judgement and have published many papers on the topic. I highly recommend studying their work if you are in a position either of having to present probabilistic risks to others or of having to evaluate such risks presented to you. --Alan Wexelblat, Reality Hacker and Cyberspace Bard,, Media Lab - Advanced Human Interface Group email@example.com 617-258-9168 firstname.lastname@example.org
> Any comment on how to communicate risk information so that people get a correct > understanding, especially when you are informing people about very small risks? Perhaps a good way to communicate information about risks is to tell people the odds and compare them with the odds of various every-day disasters. For example (off the top of my head): Odds of dying in a car accident in the next year: 1 in 10000 Odds of a mother dying during childbirth: 1 in 10000 Odds of middle-aged adult dying in the next year: 1 in 1000 Odds of a young child dying in the next year: 1 in 100 Odds of a baby dying during childbirth: 1 in 100 People are usually given insufficient credit for being able to compare risks themselves, perhaps because they are seldom handed straight numbers by experts. Stuart Wray, Olivetti Research Limited, 24a Trumpington Street, Cambridge CB2 1QA ENGLAND email@example.com +44 223 343204 fax: +44 223 313542
Xavier Xantico asks how to communicate risk information to the public in a meaningful manner. Suggestion: it would be nice to see the "automobile accident" and the "cigarette smoking" standards. Examples (purely fictional): "The risk of your child seriously injuring himself in 3 hours of playing with this toy is about the same as that of being an automobile passenger for 4 minutes." "The risk of long-term liver damage from this medication is approximately the same as the risk of cancer from smoking 2 packs of cigarettes." I realize that there would be serious problems in accurately calibrating such "standards," but I think they would be very valuable in giving people a clear understanding of risks. Robert D. Cameron, School of Computing Science, Simon Fraser University Burnaby, B.C., V5A 1S6 firstname.lastname@example.org
In a previous article, Xavier Xantico (?) writes about the difficulty of communicating statistical risks to ordinary people. The Richter scale seems to be successful in allowing people to think about and compare seismic events (earthquakes); perhaps we can develop a similar scale for statistical risks. I suggest a logarithmic (base 10) scale based on the expected lifetime frequency of the event (and population) in question. To avoid confusion, a "lifetime" should be defined as a fixed duration, 80 years perhaps, since many interesting events will be fatal. Thus, an event scoring 0 on this scale would be expected to happen once per lifetime (per person), an event scoring +1 ten times, etc. The event "dying in an airline crash" might score -6, "dying of cancer" -1, "hearing a campaign promise" +3, and so on. The approximate useful range of the scale would be -14 (a 1 in 5000 chance of occurring once in the lifetimes of 5 billion people) to +11 (expected to happen tens of times per second for each individual). One or two canonical events could be chosen for each integer on the scale to make the scale easily accessible to nonstatisticians (such as myself). --Mike Coleman (email@example.com), Lord High Executioner of Anhedonia--
> Any comment on how to communicate risk information so that people get a > correct understanding, especially when you are informing people about very > small risks? Well, what you *don't* do is say: "Every time you do X, it knocks n days off your life!" Such statements are hopelessly misleading to anyone not familiar with the concepts of "distribution of time to failure" and "mean time to failure". It is unfortunate that many government health education statements are couched in precisely those terms. Other than that, you must get across a few of the basic ideas of probability theory. The "urn" model is quite intuitive: a 0.01 probability is the probability of pulling the black ball out of an urn containing 99 white balls and one black ball. The best introduction to probability for the non-mathematical reader that I have read is "How to take a chance" by Darrel Huff (perhaps better known for his other popular book "How to lie with statistics"). Also try: "Making Decisions", D.V. Lindley, John Wiley & Sons, 2nd Ed., 1985 Lindley is a bit more demanding than Huff. (Also Huff teaches you how to play good poker! :-) To see if the ideas have been understood, ask the "student" what is wrong with the following: A businessman, who needs to fly a lot in order to do his job, is terribly afraid that there will be a bomb on board one of his flights. He asks his friend (a probability theorist) what to do about it. "Well!" says the friend, "What do you think are the chances that there will be a bomb aboard your flight?" "Judging by recent statistics," says the worried flyer, "it could be about 1 in ten thousand." "So what would be an acceptable risk?", says the friend. "How about 1 in 100 million, which is the probability of two separate bombs being on board?" "Yes, that would be OK." "Fine! So all you have to do is, every time you board a flight, make sure you carry a bomb!" Peter Mellor, Centre for Software Reliability, City University, Northampton Sq., London EC1V 0HB, Tel: +44(0)71-477-8422, JANET: firstname.lastname@example.org
One thing you have to remember is that the risk of something happening isn't the only factor to be considered when deciding whether or not to accept that risk. The consequences of the event, its preventability, and the likely consequences of alternative actions also have to be considered. To return to your example, I have a fungal infection in the nail beds on one of my feet. The only cure is to take tablets that ensure there is a fungicide in your blood, and hence fungicide reaching the nailbed. Now this drug occasionally destroys the liver, and doesn't always work. Without a liver I would be dead. The risk is low but the benefits from the drug are minor, I can live with tickly toenails, especially as I can treat the itch. It isn't worth the risk in my opinion. If I had a mestasised cancer on the other hand and the drug of choice could destroy my liver I would take it. I once went to a talk on people's assessments of the risk of having a handicapped baby, and how likely people were to risk amniocentesis/chorionic villi testing (which has a risk of inducing a miscarriage). The lecturer was startled that the willingness to be tested varied with whether or not people already had a handicapped child, how many normal children they had, and the severity of the handicap as well as the risk that the child was handicapped. He didn't see that the greater impact the care of a handicapped child would cause of the more important it was that it didn't happen and the less important the relatively fixed "cost" of a miscarriage became. He was also horrified that people who didn't want an abortion would want to be tested. I think that makes sense. Knowing gives you time to prepare, and there may be a miscarriage (even if the latter option isn't admitted to themselves). Basically what I am saying is that people aren't misjudging risks as often as is assumed by professionals, they are taking other factors into account as well. L. Bootland, University of Edinburgh, Genetics email@example.com
To: firstname.lastname@example.org Janet, I don't normally read comp.software-eng. (It takes up enough of my time reading comp.risks and a couple of others!) However, your message was passed on to me by Dave Bolton in the Computer Science Department here. (Thanks, Dave!) I hope you don't mind, but I have sent this reply, including a copy of your original message, to comp.risks also, since it is relevant to a recent discussion there. Perhaps you could forward it to comp.software-eng as well, since I am not sure of the distribution address for that list, and I'm too lazy to look it up right now! ----- Begin Included Message ----- >Newsgroups: comp.software-eng >From: email@example.com (Janet Figueroa (x37973)) >Subject: Software Reliability - how to calculate? >Summary: Require reference for software reliability calculation >Keywords: software reliability reference calculation >Organization: Clinical Diagnostics Division, Eastman Kodak Company >Date: Tue, 17 Nov 1992 18:21:19 GMT I just had an interesting discussion with a peer who is an electrical engineer. It was about our reliability budget in which we have set a goal of a certain number of service calls per year upon the introduction of the instrument. She asked if there was a methodology followed in calculating reliability for software. One of the examples we were discussing was code that resides in a PROM. If there was an error in the software and our service representative will have to replace the PROM, how would that be looked at? I told her point blank that I was not aware of any process used to calculate software reliability. Is there one? I would be very interested to know how the reliability is measured in an application programming environment as compared to how it is determined for a device driver, for example. Any insights would be appreciated. Thank you very much Janet C. Figueroa, firstname.lastname@example.org ----- End Included Message ----- > She asked if there was a methodology followed in calculating relaibility for > software. There certainly is! Basically, you observe the software during its later test and trial phases, when it is reasonable to assume that it is being run under conditions which are a fair approximation to its eventual operational environment. You record the execution time and the failures. You diagnose the cause of each failure, and strip out from your data set solely those failures which are due to activations of *new* faults (i.e., your final data set consists of instances of detecting new faults over execution time). Unless something is wrong with the environment in which you are running your software, you will observe a decreasing rate of finding new faults over execution time. You can do a statistical analysis of the growth in reliability as you debug the software, and estimate such quantities as:- - current rate of finding new faults, - number of faults that will be found in a given period, - median time to failure, - extra testing time required to get to given target levels of any of the above. To do the analysis, you can apply various software reliability models. The estimation process involves estimating the values of the model parameters from the observed history of finding faults. Be warned! :- 1. Different reliability models sometimes give different predictions based on the same set of data. (However, there are statistical methods which can measure the bias in the predictions from each model, correct for this, and so reconcile the predictions.) 2. The whole procedure depends *crucially* upon the observations being made in the same operational environment as that in which the software will be used in service. 3. The measure of "execution time" employed must be a meaningful measure of the degree to which the software has been used, taking account of the type of software and its application. 4. Serious problems arise if very high reliability is required. (Basically, the statistical methods require a reasonably large data set of faults found in order to work with any degree of accuracy. For ultra-high levels of reliability, such as those required for safety-critical systems, the observation of even one failure during the trial period means that the software is not good enough!) Apart from the restriction regarding ultra-high reliability requirements, the methods can be applied to any type of software, whether resident in a PROM or not. They have been applied "in anger" to all types of software where a modest failure-rate is acceptable, including operating systems and application software, and have given reasonably accurate predictions. Given that you are envisaging a situation where support representatives *will* be deployed (i.e., this system *is* going to fail, isn't it? :-) these are exactly the sort of methods you should be employing. Logistical calculations can be based on the reliability estimates. For example: How do we decide when to issue a "bug-clearance" release on a new PROM, in order to achieve the maximum cost-effectiveness of our support operation? There is a voluminous literature on all of these problems, and CSR has specialised in them for a number of years. If you require any further information, copies of papers, etc., do not hesitate to get in touch. Regards, Peter Mellor, Centre for Software Reliability, City University, Northampton Sq., London EC1V 0HB, Tel: +44(0)71-477-8422, JANET: email@example.com
Rebecca Mercuri (firstname.lastname@example.org) writes: > Many of the RISKS postings point to the inadequacy of software > engineering methodologies and practices, yet few colleges and > universities offer COMPREHENSIVE courses in SW Eng. and far fewer > REQUIRE them as part of core curricula for the next generation of EE > and CS professionals. City University now offers (starting this year, i.e., Oct. 92) a BEng in Software Engineering as a first degree. The BSc in Computer Science has for many years included both core and optional modules on Software Engineering, which are also taken by students on other degree programmes (Mathematics, Computer Engineering, Economics and Computing, etc.) Similar modules are also offered as options on several Master's courses. Having been heavily involved in teaching many of these courses, and being part of the design team for the new BEng in SW Eng., I would be extremely interested to hear the views of other RISKS folk about what should go into a "COMPREHENSIVE" course in SW Eng. Peter Mellor, Centre for Software Reliability, City University, Northampton Sq., London EC1V 0HB, Tel: +44(0)71-477-8422, JANET: email@example.com
RISKS readers may be interested to in the following seminar. Some UK readers may even be able to come along. (All welcome!) ************************************************************ * City University Computer Science Seminar * * Advanced technologies for automotive collision avoidance * * Raglan Tribe, of Lucas Automotive Ltd. * * Wednesday 25th November, 2 - 3pm, room A230 * ************************************************************ Every year, in the 12 countries of the European Community, some 55,000 people are killed in road accidents. The European vehicle manufacturers have joined together with component suppliers to share research on improved safety and traffic efficiency. A team at Lucas is collaborating with Jaguar to develop collision avoidance systems suitable for fitting to the majority of road vehicles. Basically, the system will use various sensing methods — optical, infrared, and radar — to detect where the road is, and the location of other objects (vehicles, people, stationary objects) that are likely to cross the path of the vehicle equipped with the system. The system then assess what threat these objects pose, and assists the driver in selecting the best course of action. Enormous processing power is necessary to extract edges in the video image, before detecting vehicles and road or lane edges. Some 3000 million operations per second must be carried out for real-time working. The next stage of the project will investigate the fusion of microwave and infrared sensors to provide a more reliable view of the world around the vehicle. All are welcome to come to this seminar on Wednesday the 25th November, from 2 - 3 pm in A230. We are very fortunate to be able to listen to Raglan Tribe, of Lucas Automotive Ltd., who will also be presenting a short video of the system. Please contact Geoff Dowling on 071 477 8442, or e-mail firstname.lastname@example.org to confirm these arrangements in case of last minute changes. Peter Mellor, Centre for Software Reliability, City University, Northampton Sq., London EC1V 0HB, Tel: +44(0)71-477-8422, JANET: email@example.com
I am looking for a PhD program in Computer Science or in Technology and Public Policy programs in the US or Canada in which one could specialize in the topics familiar to RISKS readers. So far, I have only heard of the Computing, Organization, Policy and Society at UC Irvine. Dr. Rob Kling heads a program on RISKS and related issues. It sounds good. Any others out there?
Dr. Stephen Lukasik has agreed to act as guest editor of a special issue of "The Information Society" journal addressing errors in large databases and their social implications. Attached is a brief prospectus for this special issue. If you know of data or research reports relevant to the topics mentioned in this prospectus, or can suggest relevant authors or are interested in submitting a manuscript yourself, please contact one of us. Also, I would appreciate your forwarding or posting this message to others that may have interest in this topic. If anyone receiving this message is unfamiliar with the journal and its focus and interests, I would be happy to supply additional information. (For those on the RAND mailing list: It is available for routing or perusal in the RAND library.) Thanks for your assistance. Bob Anderson - - - - - - - - - - - - - - - - - - - - - - - - - - - ERRORS IN LARGE DATABASES AND THEIR SOCIAL IMPLICATIONS Prospectus for a special issue of the Information Society Journal With the growth of information technology over time, we are becoming increasingly affected by data in electronic databases. The social and business premise is that electronic databases improve productivity and quality of life. The dark side of all this is that these databases contain errors, most trivial but in some cases they contain errors that by their nature impose a penalty on society. The penalties can range from minor annoyance and modest administrative cost in having a record corrected, to more serious cases where more costly consequences ensue, to conceivably, loss of life or major loss of property. The consequences to society of errors in electronic databases can be expected to increase, probably at an increasing rate. Some factors contributing to this expected increase are the increasing extent, in both size and coverage, of existing databases; increasing capture of data by automated transaction systems, from text and image scanners and the like; greater coupling of databases, either by administrative agreements or by more sophisticated search processes; more "amateur" database administration with increasingly widespread use of information technology; increasing use of heuristic search techniques that lack "common sense;" and probably other well-meaning but pernicious influences. The purpose of the proposed issue is to accomplish the following: (a) increase recognition of, and awareness of the growing nature of the problem of errors in electronic databases that are increasingly becoming regulators of modern life; (b) encourage greater attention to the collection of error rate data and to quantitatively assessing the social and economic costs deriving from those errors; (c) foster theoretical and empirical studies of the propagation of errors through the coupling of, or joint search of, multiple databases; and (d) encourage the formulation of measures, in both technology and policy domains, designed to limit the costs accruing from the inherent growth in size and connectivity of electronic databases. We seek papers for the issue that will focus on the following aspects of the problem addressed here: (1) an enumeration of socially relevant databases, whose errors can have important consequences, either from a large number of small unit cost consequences or a small number of high cost consequences; (2) quantitative data on errors in databases, classified by the nature of the errors and their derivative costs; (3) obstacles to a full and open discussion of the problem such as those deriving from concern over legal liability and loss of business from "owning up" to the problem; and (4) proposals for technical and policy measures that can limit the growth of the problems addressed. The premises of the journal issue are: (1) that the problems of errors in databases can not be minimized until they are adequately recognized and fixes explored by the professionals in the field; and (2) that we must move from the anecdotal level, where horror stories abound, to a quantitative level where the economics of fixes, either in quality control at the point of data collection, or the quality control of the output of database searches, can be sensibly analyzed. Your interest in contributing to this special issue is invited. Suggestions for possible topics, authors, or an interest in contributing should be communicated to one of: guest editor: editor-in-chief: Dr. Stephen Lukasik Dr. Robert H. Anderson 1714 Stone Canyon Road RAND, P.O. Box 2138 Los Angeles CA 90077 Santa Monica CA 90407-2138 net: firstname.lastname@example.org net: email@example.com tel: (310) 472-4387 tel: (310) 393-0411 x7597 fax: (310) 472-0019 fax: (310) 393-4818
I am posting this for Robert Ratner, Ratner Associates Inc, which does international consulting in air-traffic control and aviation safety issues. He is looking for a public-accessible data base on software-related incidents in this area. Email correspondence can be sent to me at firstname.lastname@example.org. Thanks. Dave "Van Damme" Ratner email@example.com [I suppose he should be reading RISKS! PGN]
Please report problems with the web pages to the maintainer