Generative Adversarial Networks GAN
In order to understand and creatively adapt Machine Learning technologies, we attempt to train our own models with data owned by Rockwell Group.
We gathered 16,000 unique variations of Floor Plans from the Rockwell Group network to train our own model.
This is the first step to developing more sophisticated tooling. This exercise is a proof of concept that opens possibilities for other forms of experimentation with Machine Learning.
Related: Floorplan placement
We trained our own model on 16,000 floor plan variants over 20+ years of RWG portfolio.
What were the limitations?
We need a set ~100,000 plans to get more meaningful information.
What does this help us do next?
Form Finding – defining the shell, filling the site
Inspiration – Rorschach Drawing, Ideation
Content – Novel Approaches to Content
Expanding the training set – renderings and 3D models
Training other types of Custom models – sensor data, camera feeds, content references