other considerations

• constrained to unconstrained

  • no complications

• global optimization

  • multiple starting points
  • ||F(t)-F(s)|| < sqrt(2E)||t -s|| for saving
  • the other side of the saddle points

• bootstrap method

  • reduce complexity + add robustness

• robust norms (M-statistics)

• exact, fast & global linear search

• cheap gradient evaluations

• CG method & diagonal pre-condition

  • VPM once in a while: LSQR

• sparse matrix methods

Previous slide

Next slide

Back to first slide

View graphic version