Design without Precedents

New methods of automating design generation differ greatly from the ways AI researchers have attempted to elicitate and implement design knowledge. The latter tradition is coined by the so called 'knowledge based approaches'. A good example for floor plan design can be found in (Zandi 1992). Using iterative algorithms (e.g. artificial neural networks), this programming work can be relieved, carrying out learning of patterns. During the learning process the parameters of the iterative algorithm are gradually adjusted until a particular input produces the desired output. Advantages of this approach can be found in the non-necessity to specify the knowledge to be implemented, or how a problem is solved. On the other hand, the knowledge captured in a neural network is not explicit and therefore not explicable. For floor plan design this approach was taken by (Zandi and Koppelaar 1992). Disadvantage of both approaches is that they both are specific for the problem at hand and it is difficult to adapt them to a wide range of varying problems. Each set of expert rules is semantically bound. Each set of learned connections in a neural net is implicitly specified by its learning set. To pursue automation of design tasks in a wide range of domains, an approach of evolving software is proposed here. The software starts from a local starting point and evolves, without any precedent, to a design. In the initial application explained in this paper, design of a floor plan is studied. The software breeds from an atomic locational egg, with a seed operating upon the egg. The 'fittest' floor plan proposals in the breed—in our case, the ones which come closest to the pre-set privacy and community design goal—are allowed to 'survive', 'breed' and 'spawn' the next generation. From a starting population of one single location it takes over a number of generations to automatically design this floor plan: