Reinforcement Learning with Transfer Learning for Cross-Context Building Design Optimization

Piragash Manmatharasan, Katarina Grolinger, Girma Bitsuamlak

Keywords: Machine Learning, ICML, Building Sustainability, Energy, Surrogate Model

TL;DR: We propose a reinforcement learning–based transfer learning framework that reuses actor–critic policies across climates and building types to enable scalable, simulation-efficient building design optimization.

Abstract: Building design optimization often relies on physics-based simulation tools, which have a high computational cost. Surrogate-assisted optimization offers a more efficient alternative by approximating simulation outputs and is typically combined with conventional optimization algorithms such as genetic algorithms or particle swarm optimization. These optimization algorithms initiate their search without any prior knowledge, requiring optimization from scratch for each new building or weather scenario. This study proposes a Reinforcement Learning (RL) approach that incorporates transfer learning through actor–critic policy reuse, enabling adaptation to new weather conditions and building types. Experimental results demonstrate improved sample efficiency, faster convergence, and reduced training variability. These findings highlight the promise of RL-based transfer learning for scalable and sustainable building design optimization.

Submission Number: 22