Evolving Nonlinear Multigrid Methods With Grammar-Guided Genetic Programming

Abstract

We formulate a formal grammar to generate Full Approximation Scheme multigrid solvers. Then, using Grammar-Guided Genetic Programming we perform a multiobjective optimization to find optimal instances of such solvers for a given nonlinear system of equations. This approach is evaluated for a two-dimensional Poisson problem with added nonlinearities. We observe that the evolved solvers outperform the baseline methods by having a faster runtime and a higher convergence rate.