Schiller, Eric2020-08-272018-052020-03-032018Schiller, Eric. 2018. Creating Novel Architectural Layouts With Generative Adversarial Networks. Master's thesis, Harvard Extension School.https://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37364552Although deep learning has made significant advances over the past 10 years, much of the focus has gone towards discriminator networks, allowing the mapping of complex input to output representing a series of classes (Goodfellow, 2014). Generative models for neural networks, on the other hand, are still in their infancy. In recent years, two particularly interesting generative models have emerged: Generative Adversarial Networks (GANs), and variational auto-encoders (VAEs). At a very high-level, a GAN is a pair of artificial neural networks that work in opposition to generate data. In this project, GANs were utilized to solve a practical problem. We designed and implemented a GAN to generate novel, two-dimensional floor plans for homes. Applications for such a generative model could include generating realistic residential neighborhoods in simulations, procedural content for games, or even providing basic prototypes for architects or designers. We also developed a floorplan dataset to train the neural network. Tools were developed in order to streamline the addition of new data to the dataset, and to allow the neural network to work with the training data.application/pdfgenerative neural networksconvolutional networksThesis or Dissertation2020-08-27