Posted by Richard Phillips.
The theme was Generative Design and how Rafal applied it to civil engineering applications.
Design representations were discussed. How generative designs differ from the traditional approach. In traditional design there is effectively a complete description of the design.
But in generative design there is no complete description but a core initial design and a generative process.
Designs are grown!
His goal is not just finding new design, but optimal designs. So he needs a fitness measure. He uses both evolutionary search and exhaustive search.
His first example was building structures, coming out of this NKS Summer School 2003 project.
Next he showed an example of traffic control system design in urban areas. One point is that you might only have local information to feed into the control method.
Rafal developed a tool called Emergent Designer to do these things, with many nice features.
For building design he did exhaustive search of the Elementary CAs. And also evolved 1D rules. For traffic control he used evolutionary algorithms on the larger 2D rule spaces. He covered a bunch of details of the experiments.
During the results section he showed some of the best designs. It's quite time intensive to get them. Compared randomly generated designs to CA generated designs, which was interesting.
A broad claim was made that for the building structures problem the cellular automata tended to make better than random designs. Certainly they seemed to have less variance if I understood the graph correctly. Also that generative designs were 6-20% better than traditional representations, and that they were qualitatively different.
He said that they are optimized well by evolutionary algorithms.
In the Q&A Rafal made the point that optimizing the CA rule and initial conditions using the genetic algorithm was better than optimizing the raw building description, since it was a small rule space that maintained (by definition) more order.
Comments