Zero-shot LQR
Preprint. One reward-free dataset, then any linear-quadratic task in closed form.
Zero-Shot Linear Quadratic Regulation from Reward-Free Data: Structured Successor Features with Stability and Finite-Sample Guarantees. Preprint.
Quang Huy Dao and Khanh Thai.
One reward-free dataset is compressed once. Any linear-quadratic task is then read out in closed form, with stability and finite-sample guarantees, and without another pass over the data.
