Minimization routine lacks constraints

[jackdawjackdaw]

The minimization routine used to find the optimal hyperparameters is currently rather basic. The pthread wrappers and driver routines are rather convoluted and could be tidied up, this is probably a secondary issue :)

My primary concern is that the actual minimizer doesn't include constraints. The user may have information about acceptable ranges for the various length scales. These should strictly be included as priors on the likelihood that is being minimized, this may require some extensive modifications to the code. The simpler case is to actually respect the ranges specified in the optstruct, currently these ranges are used when sampling initial starting points for the minimization but are not used to guide the evolution to keep it inside these ranges.

I tried a branch where the evolution is directly halted if it escapes from the allowed ranges, this is a very poor strategy rather like diving out of a car every time it looks like you won't make a corner instead of applying the breaks.

Anyone have any experience with constrained L2 minimizers? We only need simple inequality constraints: x < 3 or y > 5, as opposed to complex f(x,y,z,w) == 0 in allowed region type constraints.