Open joseph-hellerstein opened 3 years ago
One strategy is to use Latin Hypercube Sampling (LHS) to sample the parameters used for input. I've used the pyDOE implementation in the past. Also I remember a conversation with Jens Timmer a few years back who suggested the fitting itself should be done in log space and converted back at the end.
Consider multiple starting points for the optimization to have a better chance of discovering the global minimum.