A method is described for unconstrained function minimization using function values and no derivatives. A quadratic model of the function is formed by interpolation to points in a table of function values. The quadratic model (not necessarily positive definite) is minimized over a constraining region of validity to locate the next trial point. The points of interpolation are chosen from a data table containing function values at an initial grid and at subsequent trial points. The method is efficient in its use of function evaluations, but expensive in computation required to choose new trial points.