The Nonlinear Programming Solver

Multistart

Frequently, nonlinear optimization problems contain many local minima because the objective or the constraints are nonconvex functions. The quality of different local minima is measured by the objective value achieved at those points. For example, if $x_1^*$ and $x_2^*$ are two distinct local minima and $f(x_1^*) \le f(x_2^*)$, then $x_1^*$ is said to be of better quality than $x_2^*$. The NLP solver provides a mechanism that can locate local minima of better quality by starting the local solver multiple times from different initial points. By doing so, the local solver can converge to different local minima. The local minimum with the lowest objective value is then reported back to the user.

The multistart feature consists of two phases. In the first phase, the entire feasible region is explored by generating sample points from a uniform distribution. The aim of this phase is to place at least one sample point in the region of attraction of every local minimum. Here the region of attraction of a local minimum is defined as the set of feasible points that, when used as starting points, enable a local solver to converge to that local minimum.

During the second phase, a subset of the sample points generated in the first phase is chosen by applying a clustering technique. The goal of the clustering technique is to group the initial sample points around the local minima and allow only a single local optimization to start from each cluster or group. The clustering technique aims to reduce computation time by sparing the work of unnecessarily starting multiple local optimizations within the region of attraction of the same local minimum.

The number of starting points is critical to the time spent by the solver to find a good local minimum. You can specify the maximum number of starting points by using the MAXSTARTS= suboption. If this option is not specified, the solver determines the minimum number of starting points that can provide reasonable evidence that a good local minimum will be found.

Many optimization problems contain variables with infinite upper or lower bounds. These variables can cause the sampling procedure to generate points that are not useful for locating different local minima. The efficiency of the sampling procedure can be increased by reducing the range of these variables by using the BNDRANGE= suboption. This option forces the sampling procedure to generate points that are in a smaller interval, thereby increasing the efficiency of the solver to converge to a local optimum.

The multistart feature is compatible with the PERFORMANCE statement in the OPTMODEL procedure. See ChapterĀ 4: Shared Concepts and Topics, for more information about the PERFORMANCE statement. The multistart feature currently supports only the DETERMINISTIC value for the PARALLELMODE= option in the PERFORMANCE statement. To ensure reproducible results, specify a nonzero value for the SEED= option.

Accessing the Starting Point That Leads to the Best Local Optimum

The starting point that leads to the best local optimum can be accessed by using the .msinit suffix in PROC OPTMODEL. In some cases, the knowledge of that starting point might be useful. For example, you can run the local solver again but this time providing as initial point the one that is stored in .msinit. This way the multistart explores a different part of the feasible region and might discover a local optimum of better quality than those found in previous runs. The use of the suffix .msinit is demonstrated in Example 10.5. For more information about suffixes in PROC OPTMODEL, see Suffixes in ChapterĀ 5: The OPTMODEL Procedure.