2604.14803

Total: 1

#1 Anderson Acceleration for Linearly Converging SQP-Type Methods [PDF] [Copy] [Kimi] [REL]

Authors: Jonathan Frey, David Kiessling, Katrin Baumgärtner, Moritz Diehl

Although Anderson acceleration (AA) is known to speed up fixed-point iterations, it is rarely applied in constrained optimization, in particular sequential quadratic programming (SQP). We show that the local convergence behavior of a general family of (inexact) SQP-type methods can benefit from AA and introduce a simple heuristic to alleviate slower convergence farther from the solution. The method is implemented in the software framework acados. Numerical examples from optimal control illustrate consistent improvements in convergence of different SQP-type methods.

Subject: Optimization and Control

Publish: 2026-04-16 09:26:55 UTC