AI-driven design exploration: use of Reinforcement learning-based recommender system for parametric design space exploration

Abstract

In the current practice of architectural design, performance analysis is an essential step that involves simulating various design options to identify the most optimal solution. However, the process can be time-consuming, especially when the design space is vast. To address this issue, designers often use optimization algorithms to find the best solution, but simulating each design option is still a significant bottleneck. Surrogate models offer a potential solution by creating a simplified model that approximates the behaviors of the actual system. This model can then be used to simulate multiple design options efficiently. While surrogate models can help speed up the performance analysis process, they still require a significant amount of data to train effectively. Additionally, optimization done with surrogate models cannot account for aesthetic preferences, which are essential for architectural design.The paper proposes a novel design framework that leverages AI and machine learning models to address the aforementioned challenges. To demonstrate the efficacy of the framework, a parametric model is developed to generate a large number of design alternatives for a multi-story office building in Seattle. Multiple design spaces of different sizes are investigated to validate the framework. The proposed framework consists of two sections. The first section involves three consecutive layers to enable faster and more accurate prediction of performance for all design alternatives. The annual energy consumption is simulated using EnergyPlus. The first step is to convert the design parameters into weighted parameters to aid the machine learning models in understanding their distinct behaviours. The number of weighted parameters is then reduced to three using different dimensionality reduction algorithms to visualize clusters in the last step. The final step involves clustering the entire design space effectively so that the performance outcome of the centroid of the cluster can be a proper representative of all other data points in that cluster. Multiple combinations of weighting parameters, dimensionality reduction methods, and clustering models are experimented with to identify the set of algorithms that can predict the performance outcome of the entire design space with the least amount of error using a smaller number of clusters. The second section of the proposed framework involves an online dashboard that enables the exploration of the design space. The dashboard includes a reinforcement learning-based recommender system that seeks to understand user preferences through interaction and recommends similar design alternatives in each iteration. The reward function of the recommender system is customized to prioritize high-performing alternatives and pull the designer's preference in that direction. The proposed framework enables designers to explore a massive design space strategically and effectively within a short amount of time.