Core Concepts
Efficiently adapt policies for uncertain linear systems using personalized Moreau envelopes.
Abstract
Introduction to the problem of policy estimation for Linear Quadratic Regulator (LQR) in uncertain systems.
Proposal of a Moreau Envelope-based surrogate LQR cost for meta-policy adaptation.
Comparison of the proposed approach with Model-Agnostic Meta-Learning (MAML) methods.
Detailed algorithm design and convergence analysis for the Moreau Envelope-based Meta Linear Quadratic Regulator (MEMLQR).
Numerical experiments showcasing the performance and adaptation capabilities of the MEMLQR algorithm.
Empirical comparison with MAML-based approaches for policy optimization in uncertain systems.
Conclusion highlighting the benefits of personalized policy gradient methods for uncertain linear systems.
Stats
Numerical results show proposed approach outperforms naive averaging of controllers.
Proposed method has better sample complexity than MAML approaches.
Quotes
"We propose a Moreau Envelope-based surrogate LQR cost for meta-policy adaptation."
"ME provides better empirical performance than MAML in multi-task setups."