Enhancing performance and generalization in dormitory optimization using deep reinforcement learning with embedded surrogate model

Abstract

The natural ventilation and daylighting of dormitories are key factors affecting student comfort and health, with early-stage design greatly impacting indoor performance. Existing optimization methods, mainly evolutionary algorithms like genetic algorithms, excel in global search but struggle with adapting to dynamic environments, handling multi-dimensional tasks, and ensuring model generalization. This study proposes a multi-objective optimization framework integrating GAN and DRL for dormitory indoor airflow and daylighting enhancement. The GAN model provides real-time predictions of global wind speed and useful daylight illuminance (UDI). By combining a GAN-based surrogate model with a DRL approach based on Deep Deterministic Policy Gradient (DDPG), the framework iteratively refines dormitory unit layouts through continuous interaction between the environment and the agent. The effectiveness and generalization of the DRL-based method were evaluated across three dormitory typologies. Results indicate that, on the test dataset, the pix2pix model achieved R² values of 0.979 and 0.988 for predicting indoor wind and lighting conditions, respectively. Compared to traditional genetic algorithms, the DRL model demonstrated superior performance in optimizing indoor environmental conditions, achieving up to a 9.33 % improvement in wind environment optimization. The pre-trained models exhibited a certain degree of generalization across the three scenarios. This approach provides valuable support for environmentally driven indoor architectural optimization.