The resulting designs could be fed through a variety of simulators, environmental, social, construction feasibility, etc.
One could also use a supervised classifier and have people rank the resulting designs. They can know what they like, but not why.
One could be cheap and use something like "The Sims" for evaluation and then optimize building that are possible to build there.
You would probably get more realistic buildings due to the different constraints on them.
I don't think the results should be taken to serious, but I imagine it serious fun to see them. xD
Credit to the author -- I think the optimized-for-windows results have been added since this was posted here.
I posted my comment after reading the article, which included optimized-for-windows results. Of course, others may have had it sitting around in a browser tab for a while…
You'll notice as you scroll down that this article _does_ vary the parameters and weights. For instance, the one that optimizes for windows tends to create interior courtyards.
I could imagine adding parameters and weights for thresholds, areas for crafts, light on two sides of rooms, rectilinearity, water shedding, etc. etc. etc.
But I love the idea that what comes out the other end is wildly weird and yet functional, and feels organic and nest-like rather than square.