I am a little bit confused about the description of this parameter in the docs.

It says: “These values will be fixed and used for any free RandomVariables that are not being optimized”.

Does this mean that these values will not be changed?

I also see that scipy.minimize standard L-BFGS-B which assumes convexity? cant the problem be non-convex depending on priors and model-composition etc.

The text is a bit confusing, but it means that of you have 2 variables in a model and ask to optimize with respect to just one of them, the other will be fixed at the start value.

For the variables being optimized (by default all the continuous variables in the model), the start is just the initial point of the optimization (x0).

I don’t know about the convexity requirement. I guess in the case the modle logp is not convex the optimizer may finish in a local maximum. You can try different starting points to see it it arrives at different conclusions.

1 Like

you are absolutely right, the text is horribly written. However, the second part “For the variables being optimized (by default all the continuous variables in the model), the start is just the initial point of the optimization (x0).” is just what i was looking for. I must have misread the description when i read it and didnt clock “not” being optimized. Thanks.

You’re welcome Do you want to suggest a more readable description on the GitHub repository?

i was referring to my own text haha. Nah the description of the param is clear and concise, my reading not so much. ty.