Deep reinforcement learning for multi-criteria optimization in BIM-supported sustainable building design
Year: 2024
Type: article
Abstract: This paper introduces a multi-criteria optimization (MCO) framework tailored for sustainable building design, leveraging a deep reinforcement learning (DRL) methodology to adjust design parameters. The framework integrates three key components: developing an assessment index system, training a deep neural network (DNN) model, and generating an optimal solution set using the deep deterministic policy gradient (DDPG) model. The DRL approach was tested on a three-story educational building in Shanghai, where it demonstrated superiority over traditional genetic algorithms by optimizing building energy consumption, carbon emissions, and indoor thermal comfort simultaneously, improving overall building performance by 13.19%. The DDPG agent autonomously learns and enhances decision-making policies through iterative interactions with a Building Information Modeling (BIM) environment combined with DNN, offering advanced data-driven decision support for sustainable building design.
Authors: Ruchit Parekh, Charles Smith, Nathan Brown
Institutions: Hofstra University, University of Colorado Boulder
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4402891770