H.B. Keller Colloquium

Monday April 19, 2021 4:00 PM

Improving Policy Learning via Programmatic Domain Knowledge

Speaker: Yisong Yue, Computing and Mathematical Sciences Dept., California Institute of Technology
Location: Online Event

This talk explores how to leverage programmatic domain knowledge to improve policy learning (which includes reinforcement & imitation learning).  I will consider two aspects.  First, how can we express policy classes using domain specific programming languages to yield interesting inductive biases that lead to sample-efficient learning while preserving flexibility and improving interpretability?  Second, building upon the data programming paradigm in supervised learning, how can we use expert-written programs as a form of auxiliary supervision to improve the reliability of policy learning?  I will present problem framings, algorithms, and experiments for two settings: efficient learning of programmatically interpretable policies, and controllable generation of behaviors.

Series Computing + Mathematical Sciences Lecture Series

Contact: Diana Bohler at 6262326138 dbohler@caltech.edu