Differentiable MPC for End-to-end Planning and Control

J. Zico Kolter, Byron Boots, Brandon Amos, Ivan Dario Jimenez Rodriguez, Jacob Sacks

Publication date: 31 October 2018

Abstract: We present foundations for using Model Predictive Control (MPC) as a differentiable policy class for reinforcement learning in continuous state and action spaces. This provides one way of leveraging and combining the advantages of model-free and model-based approaches. Specifically, we differentiate through MPC by using the KKT conditions of the convex approximation at a fixed point of the controller. Using this strategy, we are able to learn the cost and dynamics of a controller via end-to-end learning. Our experiments focus on imitation learning in the pendulum and cartpole domains, where we learn the cost and dynamics terms of an MPC policy class. We show that our MPC policies are significantly more data-efficient than a generic neural network and that our method is superior to traditional system identification in a setting where the expert is unrealizable.

Has companion code repository: https://github.com/locuslab/differentiable-mpc

This page was built for publication: Differentiable MPC for End-to-end Planning and Control

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6309045)