CMA is much better at fitting rugged landscapes than nmpfit, but since it's stochastic it sometimes wanders off in the wrong direction, especially with small population sizes. I've gotten around this in practice by starting a few runs (e.g. 5) and letting them all run for a little while, but not to convergence. Then I take the best one and restart CMA from there, running to convergence. It would be great if there was an automated Strategy to do this, similar to the TemperedStrategy for emcee.
CMA is much better at fitting rugged landscapes than nmpfit, but since it's stochastic it sometimes wanders off in the wrong direction, especially with small population sizes. I've gotten around this in practice by starting a few runs (e.g. 5) and letting them all run for a little while, but not to convergence. Then I take the best one and restart CMA from there, running to convergence. It would be great if there was an automated Strategy to do this, similar to the TemperedStrategy for emcee.