SAMOSA: A “Simple Approach to a Multi-Objective Simplex Algorithm”
SAMOSA functions as a simple, lightweight multi-objective optimizer that is based on a modified Nelder-Mead simplex algorithm for non-linear optimization. The algorithm samples the parameter space on the vertices of a “simplex” - a multi-dimensional generalization of a tetrahedron. The simplex crawls through parameter space much like an amoeba, as its vertices reflect, expand, and contract during its search.
- Multi-objective search: SAMOSA's key innovation is its ability to perform multi-objective optimization, which is unusual for this type of algorithm. It can be run in two distinct modes on multi-objective problems:
- - Mapping mode: SAMOSA explores the trade-off surface to map its basic structure -
obviously not as thoroughly as Ferret or Locust, but well enough for some
- - Polisher mode: SAMOSA makes no attempt to map the trade-off surface in polisher mode, but rather targets the nearest part of the
surface from its starting point and returns a single solution. This capability makes it ideal for use as a precision “solution polisher” by
Ferret or Locust.
Seamless integration: SAMOSA works seamlessly with the other Qubist optimizers and understands their setup files and fitness functions. It can be called as a polisher, or even as an “inner loop” optimizer from inside of a Ferret or Locust fitness function. SAMOSA can run silently in the background or as stand-alone optimizer with its own simple interface.