You never actually told the machine that you were interested in lepidoptera, but the machine is finding out—from experience. It contains, that is, a “learning model” which stores, measures, sorts and computes the probabilities of your interests, reactions and ways of thinking. It is learning about you all right, and will soon be giving you extra information about butterflies.
Stafford Beer, “Cybernetic Thrills and Threats”
For a machine to present uninvited comments upon the qualities of a design may seem presumptuous. Yet consider that these observations might well fall into the category of “If I had only thought of…,” and so forth. Furthermore, in an evolutionary system any continual and machine-initiated surveillance would be guided by a joint maturing of the architect’s ideas along with machine observation of his methods, problems, and intents.
You are designing a soap tray, for example. Sitting at your graphic terminal with your machine, you draw an open rectangular box and specify that it is to be formed from a continuous sheet of moldable plastic. All of a sudden a bell rings or a voice speaks or some text appears on the television screen, bringing to your attention the lack of any drainage facility. How did the machine know enough to make the observation?
There are three sources for such unsolicited comments. First, you could previously have stated very specifically that all soap trays must drain water. The criterion is specific. The machine implicitly applies this maxim to its observation of your soap-tray. In this case the machine’s notice is simple and unsolicited only in time, not content.
A second way, at the other extreme of complexity, is through direct experience and real-world observation. For example, a robot might have seen bathrooms, observed soap being used, or fumbled with soap trays on its own. Such a machine might witness soap melting in water and from that make the necessary chain of observations to assume that … and so on. Even though this type of learning exceeds the scope in time of our interests, it is important that learning through groping not be underplayed or ignored; in the distant future that is how machines will probably do their learning.
A third method, more realizable in the near future, is through deduction. For example, in describing the function and the environment of the soap dish, you might have stated that soap melts in water and water runs downward. The machine, with the knowledge of the tray’s geometry and the surrounding activities, could deduce that water would indeed collect in the same place as the soap. And, since soap melts, a conflict would exist; either the soap or the water must go elsewhere.
Such machine scrutiny is particularly interesting. The facts used to deduce that the collection of water was a conflict are not necessarily unique to the design of soap trays. Water collection is a problem with roofs, sills, pavements, and so forth. In other words, after a few years of evolutionary dialoguing, a designer and a machine can establish a large repertoire of low-level axioms from which the machine can temporally deduce high-level conflicts.
But now the question arises: Why must each architect struggle with indisputable facts? He should not. Simple events—water runs downward, the sun rises in the east—would be built into the machine’s design pedigree. Their combination and association, however, must be unformed at the onset and must mature through deducing conflicts in the course of a partnership. In other words, a built-in knowledge may exist that, for example, children do not always look where they are going, and cars can kill. However, the constraint that children must not cross roads alone to get to nursery school would not be an embedded maxim.
Given a set of axioms and a set of deductive procedures, how does a machine establish the timeliness of an observation? Through context. Three types of context are particularly important: an activity context, a time context, and a rate context. Each involves ubiquitous monitoring and observing. We must assume that the machine continually tracks what the designer has been and is doing.
An activity context is the easiest to implement. Here the machine must balance between commenting on apples when the designer is working with pears. Only when the circulation pattern has been ignored by the heating system, for example, would the machine comment, directing the designer’s attention from a context of environmental to circulatory problems.
A time context is a chronology of events, a chronology of design development and design procedures. For example, the level of detail is time-contextual. A comment on bending moments is probably inopportune at the early stages of design. Similarly, in another time context, it might be more appropriate for the machine to withhold a disastrous conflict until after a weekend.
A rate context is a fine-grain time scale. It may be the most important of the three. Observation and recognition of work rates could attempt to rhythmize the dialogue. The machine would try to enter a time phasing personal to and compatible with the designer. Some people, in moments of deliberation, might enjoy a barrage of compliments and comments; others might demand complete silence. Moreover, this attitude may change with mood. Machines must discern such moods. A temporal, unsolicited comment, deduced and timely, could be antagonistic. Such prodding, however, dispels complacency and begins to transform machine servants into machine partners.