Safe Deep Reinforcement Learning for Energy-Efficient HVAC Control in Multi-Zone Residential Buildings

What the math gives to ML

The paper provides a transferable mechanism: certify that a neural policy preserves a safe state set under bounded disturbances by combining Lipschitz bounds for the policy and plant dynamics. Its key asset is a computable forward-invariance margin, rather than only an empirical constraint-violation rate. This can be transferred to RL controllers and learned dynamical systems by estimating local transition gains and constraining the policy's spectral norm so perturbations cannot cross safety boundaries.