Machine Learning for Multi-stage Selection of Numerical Methods

摘要: In various areas of numerical analysis, there are several possible algorithms for solving a problem. such cases, each method potentially solves the problem, but runtimes can widely differ, and breakdown is possible. Also, typically no governing theory finding best method, or in essence uncomputable. Thus, choice optimal practice determined by experimentation ‘numerical folklore’. However, more systematic approach needed, instance since choices may need to be made dynamic context as time-evolving system. Thus we formulate this classification problem: assign problem class corresponding that What makes an interesting Machine Learning, large number classes, their relationships. A combination (at least) preconditioner iterative scheme, making total methods product these individual cardinalities. Since very number, want exploit structure set find way classify components separately. We have developed techniques multi-stage recommendations, using automatic recognition super-clases. These shown pay off well our application area linear system solvers. present basic concepts recommendation strategy, give overview software libraries make up Salsa (Self-Adapting Large-scale Solver Architecture) project.