1: \begin{abstract}

2: We present a new algorithm to solve min-max or min-min problems out of

3: the convex world. We use rigidity assumptions, ubiquitous in learning,

4: making our method applicable to many optimization problems. Our approach

5: takes advantage of hidden regularity properties and allows us to devise

6: a simple algorithm of ridge type. An original feature of our method is

7: to come with automatic step size adaptation which departs from the usual

8: overly cautious backtracking methods. In a general framework, we provide

9: convergence theoretical guarantees and rates. We apply our findings on

10: simple GAN problems obtaining promising numerical results.

11: \end{abstract}

12: