Note on optimizers

Currently LikelihoodProfiler relies on NLopt.

Internally LikelihoodProfiler utilizes :LN_AUGLAG algorithm from NLopt to construct an augmented objective function with a set of bound constrains. Then the augmented objective function is passed to an optimization algorithm, which is defined by the keyword argument local_alg (see LikelihoodProfiler.get_interval).

Default local_alg = :LN_NELDERMEAD, which is the most reliable for the current problem among the derivative-free algorithms. Any NLopt algorithm could be potentially used as local_alg. Nevertheless we performed a series of simulation experiments displaying that some of algorithms may fail at some specific problems.

The following algorithms have shown satisfactory results on test problems (see also /test/test_grad_algs.jl):

• :LN_NELDERMEAD
• :LD_MMA
• :LD_SLSQP
• :LD_CCSAQ