FORTRAN code for minimizing a function whose uation is expensive. At each iteration, a Bayesian posterior mean for the surface shape conditional on points already sampled is constructed and the minimum of this is found. This minimum is then used as a trial point for a new function uation. A version of the program exists that takes account of the fact that the expected improvement is raised at points far points already sampled, by the fact that there is high uncertainty in such regions. There seems to be no particular performance advantage for this program over, say, quasi-newton with BFGS update. But one gets an estimate at every function uation of the shape of the function, which may be useful.- FORTRAN code for minimizing a function whose　uation is expensive. At each iteration, a Bayesian posterior mean for the surface shape conditional on points already sampled is constructed and the minimum of this is found. This minimum is then used as a trial point for a new function　uation. A version of the program exists that takes account of the fact that the expected improvement is raised at points far　points already sampled, by the fact that there is high uncertainty in such regions. There seems to be no particular performance advantage for this program over, say, quasi-newton with BFGS update. But one gets an estimate at every function　uation of the shape of the function, which may be useful.