1: \begin{abstract}

2: We consider the application of the type-I Anderson

3: acceleration (\cite{FangSecant}) to solving general non-smooth

4: fixed-point problems. By interleaving with safe-guarding steps, and

5: employing a Powell-type regularization and a re-start checking for

6: strong linear independence of the updates, we propose the first globally

7: convergent variant of Anderson acceleration assuming only that the

8: fixed-point iteration is non-expansive. We show by extensive numerical

9: experiments that many first order algorithms can be improved,

10: especially in their terminal convergence,

11: with the proposed algorithm.  Our proposed method

12: of acceleration is being implemented in 

13: SCS 2.0 \cite{SCS-2}, one of the default solvers used in

14: the convex optimization parser-solver CVXPY 1.0 \cite{cvxpy_rewriting}. 

15: \end{abstract}

16: