Bayesian Kernel Shaping for Learning Control

Poster W58

Jo-Anne Ting1, Mrinal Kalakrishnan1, Sethu Vijayakumar2, Stefan Schaal1 1University of Southern California, 2University of Edinburgh
Formulation of nonparametric locally weighted regression that estimates bandwidth (h) & regression coefficient (b) Properties: · computationally efficient · no sampling needed · automatically rejects outliers · only one prior to be specified Can be used in nonlinear methods (e.g., Gaussian processes)

Useful for computationally efficient function approximation & highly accurate local linearizations (e.g., for deriving controllers) Learnt weighting kernels at query points xq