The usual approach is to make a wild guess as to the magnitudes of the parameters, and hope for the best.
A more sophisticated approach is to use the ga (genetic algorithm) function to search the parameter space and come up with a reasonable approximation of the correct parameters. Then use fitnlm to get the most accurate set of parameter estimates, using the ga results as initial parameter estimates.
