Abstract
This dissertation presents a statistical approach to assist intelligent decision-making in the early stages of design, when coupling engineering simulation with computational design systems. The coupling of design and analysis facilitates the navigation of a design space with respect to measurable engineering criteria. In practice, this typically involves manipulating the values of the input parameters and observing the simulation output, in a trial and error fashion or automatically, using an optimization algorithm. However, typical design-analysis systems are unidirectional. Thus, when considering many parameters, it becomes very difficult to maintain control over the computational system because a cyclic approach to design and analysis externalises knowledge accumulation to human cognition. In other words, it becomes challenging to keep track of all cause-effect relationships, to generate a comprehensive understanding of the multi-dimensional design space. Instead, the research presents a statistical approach to generate abstractions of typical coupled systems in the form of probabilistic models, using Bayesian networks (BN). BNs facilitate reasoning about cause and effect relationships over multiple dimensions while their probabilistic representation, facilitates a broader representation of the design space than, with typical coupled systems. In our approach we take advantage of the fact that BNs do not distinguish between inputs and outputs and thus, enable inverse reasoning from effect to cause. The capacity to reason about inverse scenarios within a probabilistic representation enables to fix a target on a response of interest and quickly identify the input ranges that are likely to generate favourable responses within the set target. In other words, we can narrow down a vague understanding of a design space into the meaningful regions of interest. This suggests a shift from decision-making with discrete choices towards a ‘softer’ abstraction of the design space to assist the intuition. Through a case study we demonstrate how a probabilistic bi-directional mechanism can be useful for both architects and engineers to translate engineering constraints into architectural constraints and to communicate expert feedback to the architecture team in the form of soft knowledge.