Martin Roelfs

Using `CallableNumericalModel ` you could write any pythonic function you might need. For example: ```python def constraint(u_2, h): """ Do your magic here """ return np.sum(u_2 < h) - u_2.size...

Yeah I think that is a fundamental problem with the design of MINPACK (or the scipy wrapping of it, don't know at what level this occurs.) Basically, the scipy API...

There are some interesting points here. Indeed, For GradientModel we still evaluate the Hessian, but using the jacobian approximation (J^T . J). This is significantly faster than analytically calculating it...

Hmmm, well maybe you can turn it into subproblems using `CallableNumericalModel`? Or perhaps using IndexedVariables? Those already work, but there are no IndexedParameters and I guess you need those. A...

You are right in pointing out that interactive fit probably doesn't play nice with notebooks, it was designed before we really knew or cared about notebooks. Any improvements to it...

The basic ODE branch has now been merged into master because the basic functionality is there, now lets get feedback from actual use cases. All of the above points still...

> Is there a reason to want `Symbol` behaviour over `Dummy` behaviour? This is a very good question. I believe that if we coded everything correctly, there shouldn't be any...

I also think that is bad practice, and it is much easier when coding to reuse the same symbol then to type `Symbol('x')` again. But I think introducing a `DummyParameter`...

You're talking about something else: indeed LeastSquares is _also_ an objective. What I'm referring to here is a symbolic representation of the chi-squared of a model. The need for such...

Don't worry, we are clearly not talking about the same thing. I'm not even close to considering a change in the current structure. The methods `Model.chi_squared`, `Model.chi_squared_jacobian` etc., are indeed...