Search Methods
       Adaptive
           Converging Set Methods
                Controlled Random Search

Control Random Search (CRS) starts with an initial set S of points sampled uniformly in A. Iteratively new trial points are then generated and replace the worst point W in S if they are better. The algorithm stops when all function values are close enough. ([Price 1977], influenced by [Becker and Lago 1970])

In one iteration trial points are generated based on a random choice of points r₁,...,r_n+1 from S. The centroid G of the points r₁,...,r_n is calculated. The image point P of the pole r_n+1 with respect to G and the point Q halfway between the pole and G are determined.

If the f(P) < f(W) than P replaces W in S, otherwise if f(Q) < f(W) and the failure rate is more than 50 % then Q replaces W in S or else a failure is recorded.

Several modifications of Price's original algorithm, described above, exist. They differ in the choice of the subset r, how the trial points are chosen, and use of local minimization. A resent modification uses quadratic approximation instead of the original simplex method to generate trial points [Ali, Törn and Viitanen, 1997].