SPATIAL DESIGN
learning and optimisation algorithms

date: 2004 - present

The design of architecture is the design of a highly complex, organised system. This research investigates the automation of various aspects of this process with the aid of machine learning and optimisation algorithms that incorporate simulations of social interaction. The hypothesis is that knowledge gained through simulation of a system’s behaviour allows the computer to make the kinds of judgements or decisions that can be used to guide the design. The current focus is in planning of spaces to reflect given social relationships as quantified by space syntax techniques.

Two attributes can be evaluated in the comparison of spatial solutions: the behaviour of people within the space, and the various spatial qualities that inform that behaviour. Naturally there is usually a strong correlation between the two. Depthmap software will provide both visibility graph analysis to determine qualities such as spatial integration, and agent simulation to estimate the global behaviour of people within the space. This is will be treated as the spatial equivalent of the FEM analysis used in the previous research for several reasons:
-- It gives an easily quantifiable and accurate simulation of the overall system.
-- It produces this by making a discretised approximation of an essentially continuous system of movement in space.
-- The local and overall behaviour in both cases is dependent on the organised complexity of the whole system.

Various attributes are quantified by Depthmap using VGA (integration, connectivity, mean depth, entropy, etc.) and agent simulation is also possible.

As with initial structural research, two strategies exist for the definition of an objective function in spatial optimisation: a known explicit goal can be set in advance and maximised, or a set of attributes can be derived from known examples and matched by the algorithm in the new design. In both cases these must be derived from quantifiable properties of the analysis.

An initial explicit goal is suggested by space syntax studies of social behaviour in workplaces. In offices, for example, there is a positive correlation between the frequency a given person is encountered and their spatial integration (Penn, 1999), which leads to their perceived usefulness to others in the company. A clear range of differentiated to integrated space is necessary for the range of activities of the business. The case is similar regarding research laboratories (Hillier, 1996), in which it is desirable to have both high diversity of integrated spaces and high diversity of distances to different activities. The social structure of the workplace is seen as a creative system, and this measurable diversity is desirable to the creation of knowledge. In this case the objective of optimisation would be to maximise the range of integration (and circulation distances to various tasks) in the design. As with the initial structural research, the final results of this cannot yet be predicted, but it may reveal familiar typological patterns that have evolved naturally in actual spaces.

In the second strategy, no explicit judgement need be made as to what features constitute a good space, but a set of attributes can be derived from one or more examples of designs judged implicitly to be good. This is, in effect, what is happening with the integration range mentioned above, but excludes the analysis and study which explains the benefits of the attribute. A wider range of measurements can be used (anything that can be quantified by Depthmap), which will likely allow greater subtlety in the design process. This strategy corresponds closer to the embodied systems model of creativity in deriving inspiration from a domain of examples, and coincides with the next stage in the method.

This research is scheduled to continue through 2006.