曝光台 注意防骗
网曝天猫店富美金盛家居专营店坑蒙拐骗欺诈消费者
would be submitted to extensive testing and would not be used unless they demonstrated a high level of
reliability. Although the data was generated randomly, the test case generator was written for, and tailored
to this application. Once a program passed its acceptance test, it was considered complete and was entered
into the collection of versions. The acceptance test that was used represents a realistic amount of validation
for this type of software as is discussed in section seven, and resulted in highly reliable programs as is
shown below.
One result of the earlier experiment at RTI was some difficulty with machine precision differences
between versions. Although two programs computed what amounted to the same result, different orders of
computation yielded minor differences which gave the impression that one or more versions had failed. To
prevent this, all programmers in this experiment were supplied with a function to perform comparison of
real quantities with limited precision. The programmers were instructed to use this supplied function for all
real-number comparisons.
Once all the versions had passed their acceptance tests, the versions were subjected to the
experimental treatment which consisted of simulation of a production environment. A test driver was built
which generated random radar reflections and random values for all the parameters in the problem. All
twenty seven programs were executed on these test cases, and the determination of success was made by
comparing their output with a twenty-eighth version, referred to as the gold program. This program was
originally written in FORTRAN for the RTI experiment and was rewritten in Pascal for this experiment. As
part of the RTI experiment, the gold program has been subjected to several million test cases and we have
- 7 -
considerable confidence in its accuracy. It was also subjected to an extensive structured walkthrough at
UVA after translation to Pascal.
A gold version was used so that a larger number of test cases could be executed than would be
possible if manual checking of the outputs was performed. Naturally, it is possible (but very unlikely) that
a common undetected fault existed in all 28 versions, including the gold version. This would have no effect
on our final results, however, and any additional undetected common faults would only strengthen our
conclusion.
A total of one million tests were run on the twenty seven versions written for this experiment and the
gold program. Although testing was not continuous on any of the machines, a total of fifteen computers
were used in performing these tests between May and September of 1984; five Primes and a dual processor
CDC Cyber 730 at UVA, and seven VAX 11/750’s and two CDC Cyber 170’s at NASA Langley Research
Center.
3. PROGRAMMER’S BACKGROUNDS
During this experiment, the programmers were asked to maintain simple work logs, and to fill in
questionnaires about their backgrounds. In this section we give general information about the
programmers’ previous experience, education, and effort level obtained from the logs and questionnaires.
This data is provided in summary form only. We choose deliberately not to associate specific background
information with the individual versions in order to protect the identity of the programmers. This section is
based on twenty six questionnaires. The questionnaire from one programmer could not be obtained for
analysis.
Fourteen of the programmers were working on bachelors degrees and had no prior degree, eight on
masters degrees and held at least a bachelors degree, and four on doctoral degrees and held at least a
- 8 -
masters degree. Of those who held bachelors degrees, four were in mathematics, three were in computer
science, and there was one each in astronomy, biology, environmental science, management science, and
physics. All of the programmers had taken a number of undergraduate courses in mathematics and
computer science. Most had taken several graduate courses in computer science and some graduate
mathematics courses. The number of undergraduate coursework hours varied from six to 45 in computer
science, and from 12 to 45 in mathematics. The number of graduate coursework hours varied from zero to
30 in computer science, and from zero to 19 in mathematics.
The programmers’ previous work experience in the computer field varied from none at all to more
than ten years. Most programmers had only worked for a few months, usually in some form of vacation
employment.
The programmers were asked to rate their knowledge of Pascal as either expert, thorough, fair, or
limited. Of the twenty six, four rated their knowledge as expert, eighteen as thorough, and four as fair. The
中国航空网 www.aero.cn
航空翻译 www.aviation.cn
本文链接地址：
航空资料35(188)
