The Coordinate Search Solver
The Coordinate search solver aims at improving the objective function along the coordinate directions of the control parameter space. The step lengths are decreased or increased according to the values of the objective function. The Coordinate search solver does not directly evaluate gradients of the objective function. Gradients are not available for all types of parameters or might not be mathematically well defined in certain circumstances. One example is an objective function that contains noise.
However, when the solver has collected enough information around the current search point, an estimate of the gradient is constructed and a line search along this direction is attempted before a new evaluation along the coordinate directions. This accelerates the search procedure, in particular for points close to (local) minima. The algorithm is based on the description in Ch. 7 in
Ref. 1
.