The role of latent representations for design space exploration of floorplans

Abstract

Floorplans often require considering numerous factors, from the layout size to cost, numeric attributes such as room sizes, and other intrinsic properties such as connectivity between visible regions. Representing these complex factors is challenging, but doing so in a representative and efficient way can enable new modes of design exploration. Existing image and graph-based approaches of floorplans' representation often failed to consider low-level space semantics, structural features, and space utilization with respect to its future inhabitants, which are all the critical elements to analyze design layouts. We present a latent-space representation of floorplans using gated recurrent unit variational autoencoder (GRU-VAE), where floorplans are represented as attributed graphs (encoded with the abovementioned features). Two local search approaches are presented to efficiently explore the latent space for optimizing and generating new floorplans for the given environment. Semantic, structural, and visibility metrics are evaluated individually and as a combined objective for optimizations.

Keywords: Floorplan representation; GRU variational autoencoder; LSTM autoencoder; attributed graphs; floorplan generation; floorplan optimization; human behavioral features; isovists; latent search space.

Figure 1.

A sample (from the initial research) retrieval of nearest neighbors from the floorplans embedding trained using a combined set of design-semantic and behavioral features. Top-5 nearest neighbors for the queried (input) floorplan is shown. In the first row, room types and square footage areas are shown, whereas, in the second row, the time-based behavioral dynamics for human–building interactions are shown as color-coded heat maps, where red areas highlight overcrowded regions in space.

Figure 2.

Our framework consists of two phases: (1) offline phase—First, the input floorplans are converted into attributed graphs. We then generate sequences of the attributed graphs using random walks. These sequences are then fed into the embedding model to train and learn the embedding space. (2) online phase—Given an input floorplan, we convert it into an attributed graph and generate its sequences using a random walk. These sequences are then fed into our embedding space. We then call Retrieval or Generative procedures for retrieving similar nearest neighbors or generating a new graph for the input floorplan in the direction of the targeted objective, respectively: (a) floorplan to attributed graph, (b) graph to sequences, (c) generative optimization learning embeddings with a parallel GRU-VAE, (d) retrieval-based optimization, and (e) generative optimization.

Figure 3.

The images above show the four phases of floorplan generation: (a) defining the exterior, (b) creating corridors, (c) creating rooms, and (d) connecting rooms and corridors.

Figure 4.

The above figure shows 12 types of floorplans that differ based on their exterior shape (e–g) and interior arrangement of corridors (a–d).

Figure 5.

The left image shows a graph (red) overlaid on its corresponding floorplan. The right image shows the nodes’ isovists overlaid on the same floorplan. Pixels that are overlapped by many isovists are dark blue, and pixels that are covered by few isovists are light blue.

Figure 6.

Each row shows an initial graph and five graphs that improve upon an objective function using a retrieval-based approach.

Figure 7.

Each row shows an initial graph and five graphs that improve upon the minimization of one feature and maximization of another.