Conjugate grids for unconstrained optimisation
Several recent papers have proposed the use of grids for solving unconstrained optimisation problems. Grid-based methods typically generate a sequence of grid local minimisers which converge to stationary points under mild conditions. The location and number of grid local minimisers is calculated for strictly convex quadratic functions in two dimensions with certain types of grids. These calculations show it is possible to construct a grid with an arbitrary number of grid local minimisers. The furthest of these can be an arbitrary distance from the quadratic's minimiser. These results have important implications for the design of practical grid-based algorithms. Grids based on conjugate directions do not suffer from these problems. For such grids only the grid points closest ( depending on the choice of metric) to the minimiser are grid local minimisers. Furthermore, conjugate grids are reasonably stable under mild perturbations so that in practice, only approximately conjugate grids are required.
- Engineering: Reports