Online Control via Counterfactual Tracking

What the math gives to ML

The paper separates policy selection from physical execution: arbitrary causal policies are simulated counterfactually, while a fixed stabilizing feedback controller tracks their predicted state-input trajectory on the real plant. This provides a safety and robustness layer for an ensemble of neural policies without requiring shared architectures, memory lengths, or controller orders. The strongest neural-network adaptation is an online Hedge or PAC-Bayes selector over heterogeneous policies combined with a linear error-feedback tracker. It is most applicable to online reinforcement learning or model-based control with a known or locally learned linearized plant.