I was surprised that a different representation was not searched for. The spatial-partitioning used seemed to cause many problems. Maybe they could find nothing better for their purpose but it would have been good to describe the alternatives.

      The mass evaluation criteria does not specify what an example value of ideal mass would be, but rather just says one is specified. Throughout the rest of the article, a lower mass is assumed to be better than higher mass. It appears in many examples. One would think an example of needing a higher mass would have been given as well but it is not. I also found the separation of "stability" and "supportiveness and stability" counter-intuitive. The means by which the various evaluation criteria are combined to produce an overall fitness for a design is not addressed. The system must have done this somehow so it couldn't be that difficult to explain.

      In the introduction to the article there is a wonderful air of excitement. The prospect of reading about a system that claims to create entirely new designs from scratch was very inviting. Later on we read that the system must be given a number of "primitives" to start with. These primitives are actually the number of boxes that the table will be made from. The number and type of components are already chosen. It makes the design work being done seem more like configuration, and certainly not "from scratch" despite the very convincing results. It is suggested that the system can be made to evolve the number of primitives. This is not said to be a feature that helps achieve the "design from scratch" goal but as a fix for artificially limiting the range of good results the GA can produce.

      While genetic algorithms do not seem to have any inherent characteristics that lend themselves to better understanding design as a process, it can serve very well as a basis for comparison. As knowledge is identified to help guide the design process, a GA based system can be augmented with this knowledge. A comparison of performance of this system and one without that knowledge may give a measure of how useful the knowledge is. Measures of this sort have been very difficult for other systems. I was surprised the authors did not mention this point. Their work can be used to support this. One example from this article is how the system was made to produce symmetrical designs.