Dynamic Gradient Descent and Reinforcement Learning for AI-Enhanced Indoor Building Environmental Simulation

We propose a novel dynamic gradient descent (DGD) framework integrated with reinforcement learning (RL) for AI-enhanced indoor environmental simulation, addressing the limitations of static optimization in dynamic settings. The proposed method combines a hybrid optimizer—stochastic gradient descent with momentum and adaptive learning rates—with an RL-driven meta-controller to dynamically adjust hyperparameters in response to real-time environmental fluctuations. The core innovation lies in the time-varying optimization landscape, where a Transformer-based policy network modulates the learning process based on a reward signal that balances prediction accuracy and parameter stability. Furthermore, the system employs a multilayer perceptron predictor trained on computational fluid dynamics-augmented data to model nonlinear thermal–airflow interactions, replacing conventional lumped-parameter models. The integration of these components enables autonomous adaptation to short-term disturbances (e.g., occupancy changes) and long-term drifts (e.g., seasonal variations) without manual recalibration. Experiments demonstrate that the framework significantly improves simulation accuracy and control efficiency compared to existing methods. The contributions include a unified adaptive optimization-RL architecture, a closed-loop hyperparameter control mechanism, and scalable implementation on GPU-accelerated hardware. This work advances the state-of-the-art in intelligent building systems by enabling self-tuning simulations for real-world dynamic environments.