Abstract
The main objective of this research paper is to investigate the localconvergence characteristics of Model-agnostic Meta-learning (MAML) when appliedto linear system quadratic optimal control (LQR). MAML and its variations havebecome popular techniques for quickly adapting to new tasks by leveragingprevious learning knowledge in areas like regression, classification, andreinforcement learning. However, its theoretical guarantees remain unknown dueto non-convexity and its structure, making it even more challenging to ensurestability in the dynamic system setting. This study focuses on exploring MAMLin the LQR setting, providing its local convergence guarantees whilemaintaining the stability of the dynamical system. The paper also presentssimple numerical results to demonstrate the convergence properties of MAML inLQR tasks.