Unforutunately, there's currently no way to specify different initial estimates for different categorical covariate values. I can see how this might be useful for improving the optimization, so I will make sure we have a record of this in our enhancement database. If you want to share more details on why this feature would be useful, please feel free to comment here or reach out to me directly.
If this is functionality critical for you, I could suggest various (admittedly painful) workarounds. The first one I thought of was to create separate model parameters for each covariate value. Another idea is to "intercept" the call to the optimization function so that you can replace the initial estimates. I'm happy to brainstorm further if that's of interest.
-Arthur
