To some of the more hidebound architects the concept of a machine generating threedimensional solutions is immoral, impossible, or endorses unemployment for threatened architects. The premise is that a human architect’s experience gives him license to be the exclusive translator of human requirements into physical form.
Physical form, according to D’Arcy Thompson (1917), is the resolution at one instant of time of many forces that are governed by rates of change. In the urban context the complexity of these forces often surpasses human comprehension. A machine, meanwhile, could procreate forms that respond to many hereto unmanageable dynamics. Such a colleague would not be an omen of professional retirement but rather a tickler of the architect’s imagination, presenting alternatives of form possibly not visualized or not visualizable by the human designer.
An architect would not and should not confront a “criteria machine” to decrease visual privacy, increase public access, and watch contortions of form on a television screen. Instead, in the rhythm of the dialogue, a solution-generating capacity would be an evolutionary enterprise where the machine would act in “interrupt” or “reply” to its partner’s activity. The architect might search for a configuration by observing the machine’s attempts at satisfying a statement of the problem, or the machine might learn by observing the architect. In such a system both the architect and the machine could interrogate each other in order to locate those characteristics of the site and the criteria that imposed certain factors and courses of action on the generated solutions.
There are two distinct types of generated solution: one accommodates underconstrained problems; the other works within overconstrained situations. The underconstrained situation (rare in architecture) has a large set of possible solutions. The criteria are satisfied by many alternatives. These alternatives must then be evaluated by the architect using “intuitive” means, selection criteria he either does not understand or has never presented to the machine.
In the overconstrained problem, the generating mechanism is presented with great amounts of factional criteria that no form can completely satisfy. The generating mechanism searches for a solution that best relaxes the constraints, a point of greatest “happiness” and least “friction.” The resulting form is a status of criteria compromise where the constraints least antagonize one another.
Both problem types involve trial-and-error procedures, tasks well suited for self-improving machines. In many cases random numbers are employed; they deserve mention, as their use is often misunderstood. In solution generation, a random number is a substitute for missing information or unpredictable information. Rather than just cast a Monte Carlo atmosphere of surprise, random numbers simulate non-deterministic events such as family displacements, employment changes, physical expansion. Usually, as a system grows, events become more and more deterministic, and the possible alternatives diminish. Generating procedures can appropriately acknowledge this sort of growth by changing the distribution of “randomness” in response to the present state of the form, as described by previous actions, external information, and stage of growth.
As one example of solution generation, a student project—LEARN—was developed by a group of M.I.T. Master’s of Architecture students who had no previous computer programming experience. LEARN was a computer mannerist. It watched the designers’ activities by observing ten simple solutions. (In this case they were “sugar-cube” models transcribed to punch cards describing x-y-z centroid locations of solids and voids). Following these ten archetypes, the machine was asked to generate a solution of its own. The appeal of this simple experiment is that the criteria were first determined from the form and then used in the generation of the alternatives. The students observed the variations within the given “style” of the solution. The mannerism was derived from the original ten solutions and was then updated by the eleventh. The machine proceeded to generate a twelfth solution, updated its “manner,” generated a thirteenth, and so on. After a denouement of five thousand separate solutions to the same problem, the mannerist machine did not generate or embark on wild tangents. In fact, the conviction of the machine was so intense that the last thousand solutions had little distinguishing variety.
A second example is GROWTH, also a student project. This system operated within a larger work space (approximately a square mile) than LEARN and did not observe a specific designer’s methods. The generated solutions were periodic glimpses at stages of growth. The computer employed the principle of “influences,” where each element’s status (solid or void) was determined by its “conviction” (to be what it was or to be what it was not). As soon as a void became solid or a solid became void, ripples of influence would disperse, locally disturbing the convictions of adjacent elements (in proportion to proximity and activity relationships). A solid might become more convinced of its solidity or else an adjacent void might tend toward a state of solidity, being now unconvinced of its status. In effect, the rules of conviction were the generating force. For example, a lone ten-foot cube in the middle of a large field might influence its void neighbors under one set of rules to be less convinced of their voidness and accordingly raise their probability of changing state in the next stage of growth. Meanwhile, another set of rules might make the edge members of a large complex thoroughly convinced voids or thoroughly convinced solids. The same rules might tend to lower the conviction of deeply embedded solids (in order to avail the form of interior open spaces in response to size).
A third example is the ongoing research of Timothy Johnson and Richard Krauss at M.I.T., under NSF contract (T. Johnson et al., 1969). Under the direction of Albert Dietz, this space-allocation work employs sophisticated mathematics and sophisticated graphics to optimize cross-coupled constraints and display the results. The generated solutions are composed of “use-surfaces” and “use-volumes.” They are the result of optimization techniques that assume missing information (as opposed to replacing it with random numbers). A generated solution is a function of weighted proximities, orientations (site and exposure), visual access, acoustical access, circulation, and others to be implemented. It is displayed to the user for his consent and rearrangement. Subsequently, the machine regenerates a solution more specified than the last but in the same tenor as the last. Because the machine does not explore divergent tacks, it could channel the unwary user in the wrong direction.