Bayesian optimisation

[MichaelClerx]

For paper doing CMA-ES on hyperparameters, we might want to compare with "Bayesian optimisation", which has been called the gold standard for tuning [citation needed].

As was the case CMA-ES, for which Arnaud did a "plain" implementation, it might be necessary to implement one of these methods ourselves, so that we're testing the core idea behind the method, not some battle-hardened competition code.

Then again, maybe we want to save ourselves the trouble and use an industry-standard method?

See also #684