The Two-Environment Approach: The A3 Case

A computer has made history in Boston by convincing five judges that it was actually a
person.
To some, the thought of a thinking machine is a nightmare. But to the group of computer This section describes the A3 case study, a project undertaken in 1983 with an application from laboratory
buffs who gathered in Boston for the annual “machine intelligence contest”, the outcome medicine (Sandahl, 1983; Sandahl et al. 1985; Sandahl et al. 1990). The case was our first attempt to achieve
was evidence of progress in their quest for a computer with a “brain”. a methodology applicable with limited resources, still yielding a knowledge base of high quality. The
The contest pitted a panel of judges against a bank of computers. As they keyed in their implementation of a two-environment architecture will also be discussed.
questions, at least two of the computers responded for themselves. Others were controlled 2.1 The problem
by concealed humans – the judges did not know which. Five out of ten judges were fooled.
“If a machine can convince at least one judge that it was human, we would call it At the Blood Center of the Regional Hospital in Linköping, a MUMPS (Bowie and Barnett, 1976) database
crossing the line into true machine intelligence,” said the event's compere. system runs on a VAX 11/750 computer for the management of blood donors and their donations. The
This was the first public application of the classic test of computer intelligence proposed laboratory staff, in total consisting of about 50 persons with laboratory technicians in the majority, daily
by Alan Turing, the British mathematician, in 1950. Turing maintained that if a human was investigates about 200-250 blood samples, primarily for ABO-RH determination and antibody analysis
unable to tell whether responses were being generated by a computer or by another person, (Zmijewski and Fletcher, 1972).
the computer could be called a thinking machine. Only now, however, are computers The task for the A3 system (Antibody Analysis Advisor) is to guide a laboratory technician in the first step of
advanced enough to put this test into practice. the series of tests leading to the identification of suspected irregular antibodies. A correct, though not
In round one, each computer conversation was restricted to a single topic. The subject superfluous, sequence of various analysis techniques has to be chosen.
of “whimsical conversation” generated this memorable interchange:
Computer: “It is absurd to say that a man can't love a woman all the time as it is Due to the relatively short time devoted to knowledge acquisition (one month), it was almost impossible to
to say that a violinist needs several violins.” understand and implement a deep model of the problem. For such an approach we refer to work by DeJongh
Judge: “Wonderful.” and Smith (1990). Hence, a purely heuristic model had to be built, which in turn implied that a human expert
Another judge: “What is the definition of a romantic relationship?” become our single knowledge source. Another effect of the limited resources was the necessity to keep the
Computer; “I try to be agreeable.” domain limited throughout the project, even though several closely related areas could be imagined.
The computer also showed bibulous tendencies: Another problem was the difference in requirements between the development environment and the delivery
Judge: “Have you ever had dry martini?” environment regarding, for instance, integration with existing routines. The delivery system is described
Computer: “I mostly drink one before dinner, sometimes during dinner. I don't above. The development environment had to be more flexible and support incremental development of the
like martinis after dinner.” system. Among the few expert system shells that were available at the time, EMYCIN (van Melle, 1981)
Judge: “What about wine?” running on a DEC-2060 was chosen.
Computer: “Mostly I drink relatively young burgundies of modest fame such as
Savigny-les-Beaune and Pernand-Vergelesses.” 2.2 Assumptions
When discussing “grade-school topics”, some remarks were obviously generated by A feasibility study had shown that EMYCIN was not sufficient in providing support for the creation of a
non-thinking machines: transparent knowledge base where the built-in end-user interface tool could be effectively used. For this
Judge: “Do you go to a Baptist school?” reason, it seemed necessary to develop some kind of control structure on top of the EMYCIN shell; preferably
Computer: “Be careful with non-metals, they are often brittle.” this structure should be implemented in EMYCIN itself to increase transparency and maintainability. It was
It was a far cry from the exchanges predicted bu Turing. He had forecast that by the believed that the best strategy was to start with a rather advanced structure, however partially implemented,
years 2000, a human-computer conversation might go like this: and negotiate with the expert to drop features as they showed not to be used in early prototypes. Typical
Human: “In the first line if Shakespeare's sonnet which reads <<Shall I compare ingredients of such a control structure were:
thee to a summer's day?>> would not <<a spring day>> do as well, or better?”
Computer: “It wouldn't scan.” • Help parameters. These would be used for guiding the interface mechanisms, for instance, by finding
Human: “How about <<a winter's day>>? That would scan all right?” shortcuts, by preventing superfluous reasoning, withdrawing tentative conclusions, and by making the
Computer: “Yes, but nobody wants to be compared to a winter's day.” dialogue more intelligent.
• Procedural knowledge. Such knowledge could be, for instance, to guide the order in which the output is
Hugh Loebner, the businessman and philanthropist who donated £57,000 to fund the presented.
contest, said his interest was personal: “I've always been lazy and I’d like to see the
computers do all the work.” • Coding conventions. One convention might account for the design of text strings to be automatically
combined into explanations.
He disregarded criticism that the computer's conversation was limited. “What we have Owing to the different environments on the development machine, DEC-2060, and the delivery machine,
to look at isn't how bad the programmes are now, but how much improvement there will VAX 11/750, some kind of support for the migration of the resulting knowledge base was regarded as
be,” he said. necessary.