Inspired by the comments in #432: currently we pick the initial points at random, and discussed using a Sobol sequence (or other quasi random sequence) instead.

Finding a good implementation of a QR sequence generator in many dimensions isn't soo easy.

However I just realised that we pick the number of initial points "up front", which means we can design a "optimal" grid (latin hypercube?) and evaluate the objective using that. Sobol or random only wins over this static allocation if you do not know how many points you will sample. This would be great because I think latin hypercubes should be much easier to code up than a Sobol sequence.

Would be good to hear some opinions on this.