Reproduce “Decision-Focused Learning for Complex System Identification: HVAC Management System Application” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pietro Favaro, Jean-François Toubeau, François Vallée, Yury Dvorkin
Researchers from the University of Mons and Johns Hopkins University developed a Decision-Focused Learning (DFL) framework to simultaneously identify system dynamics and optimize control policies for complex, non-differentiable black-box systems, with a focus on HVAC management. The DFL model reduced electricity cost prediction error from 389€ (conventional approach) to 16€ in a multi-zone building simulation, while also achieving slightly lower actual ex-post costs.
Key insights
- Learning system dynamics in a task-aware, end-to-end manner significantly improves the accuracy of downstream decision-making and control performance compared to traditional two-stage identification and optimization.
- It is possible to apply DFL to complex, black-box systems with unknown and non-differentiable dynamics by formulating a supervised loss function based on observable system responses, effectively bypassing the need for internal system gradients.
- Robustness mechanisms like constraint relaxation and warm-starting are crucial for stabilizing DFL training when uncertain parameters are in the constraints, preventing infeasibilities and ensuring meaningful gradient updates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.