On Sun, Feb 22, 2009 at 17:56, John Travers <jtravs@gmail.com> wrote:

Hi All,

Attached is a patch which adds two new ODE solvers to the scipy.integrate.ode module. The solvers are dopri5 and dop853, which are explicit Runge-Kutta pairs originally developed by Dormand and Prince. The fortran code was downloaded from:

http://www.unige.ch/~hairer/software.html

The license is clearly BSD and SciPy compatible.

These are excellent solvers, described in detail in the authors book:

Solving Ordinary Differential Equations. Nonstiff Problems. 2nd edition. Springer Series in Comput. Math., vol. 8.

The dopri5 code is what Matlab's ode45 is based on.

I think they would be a good addition to SciPy and I have used them often (in Fortran). The attached patch tries to follow the (somewhat strange to me) coding practices of the current ode module. I have added them to the test suite, but note that there are not many tests! (I might add some if I get time). I have also tested them with my own code which uses the ode module.

If I get the go ahead from some regular SciPy contributers then I'll go ahead and commit this patch. I think I still have SVN access, but I wanted this pacth to be reviewed first as it has been a long time since I did anything with SciPy.

* Typo: "first 0rder" * You're missing dop.pyf * I don't like the conditional imports inside the classes. * Don't print things. Use warnings.warn() if you must. But otherwise, +1. -- Robert Kern "I have come to believe that the whole world is an enigma, a harmless enigma that is made terrible by our own mad attempt to interpret it as though it had an underlying truth." -- Umberto Eco

Sun, 22 Feb 2009 23:56:55 +0000, John Travers wrote:

Attached is a patch which adds two new ODE solvers to the scipy.integrate.ode module. [clip]

I think this is a good point to discuss some API design decisions on scipy.integrate.ode. There are currently two main interfaces to ODE integration: - vode: a class, requires people to repeatedly call .integrate() to get values at separated points. Parameters set via method calls to the object. Uses DVODE/ZVODE. - odeint: a function, computes values at points given as arguments. Parameters set via (keyword) arguments to the function. Uses LSODA. Clearly, here we have one interface too many, and it's a bit of a mess. Either both LSODA and DVODE should be available only via one way (or both ways, as we decided to go with scipy.interpolate). Which to deprecate? Also, how to specify the integrator to use: choose the correct function, or specify the name of the integrator as a string? I'd perhaps like to see: - All integrators moved to classes (with CamelCase names). If you want to use eg. the ZVODE solver, you'd instantiate 'Zvode' class. - The 'integrate' method would support getting multiple time points at once. - There'd also be thin wrapper functions (with lowercase names), e.g. 'zvode' that would allow all solvers to have a simple odeint-type interface. - Both of the current 'ode' and 'odeint' interfaces would be dropped in Scipy 1.0, and deprecated before that. How would this sound? -- Pauli Virtanen

Hi John,

Attached is a patch which adds two new ODE solvers to the scipy.integrate.ode module. The solvers are dopri5 and dop853, which are explicit Runge-Kutta pairs originally developed by Dormand and Prince. The fortran code was downloaded from:

http://www.unige.ch/~hairer/software.html

This is good news, and the scipy module certainly needs an updated API. I hope that previous discussions on this list about API changes will be looked up as there were some good suggestions then. I wonder how much extra work it would be to include H&W's stiff and delayed ODE and DAE solvers such as Radau, Retard, and Hem? Those would be of great value to Scipy users, I think, as there's little high-level language support available for those AFAIK (Radau is in PyDSTool but not the others). Thanks, Rob