Successive Convexification with Feasibility Guarantee via Augmented Lagrangian for Non-Convex Optimal Control Problems

TL;DR

This paper introduces an enhanced successive convexification algorithm for non-convex optimal control problems, incorporating augmented Lagrangian methods to guarantee feasibility and convergence to a local optimum.

Contribution

It extends the SCvx algorithm by integrating augmented Lagrangian techniques, ensuring feasibility and convergence guarantees for non-convex optimal control problems.

Findings

The proposed algorithm guarantees convergence to a feasible local solution.

Numerical examples demonstrate improved convergence properties.

The method retains the favorable features of the original SCvx algorithm.

Abstract

This paper proposes a new algorithm that solves non-convex optimal control problems with a theoretical guarantee for global convergence to a feasible local solution of the original problem. The proposed algorithm extends the recently proposed successive convexification (SCvx) algorithm by addressing one of its key limitations, that is, the converged solution is not guaranteed to be feasible to the original non-convex problem. The main idea behind the proposed algorithm is to incorporate the SCvx-based iteration into an algorithmic framework based on the augmented Lagrangian method to enable the feasibility guarantee while retaining favorable properties of SCvx. Unlike the original SCvx, this approach iterates on both of the optimization variables and the Lagrange multipliers, which facilitates the feasibility guarantee as well as efficient convergence, in a spirit similar to the alternating direction method of multipliers (ADMM) for large-scale convex programming. Convergence analysis shows the proposed algorithm's strong global convergence to a feasible local optimum of the original problem and its convergence rate. These theoretical results are demonstrated via numerical examples with comparison against the original SCvx algorithm.