You are here

Running models : Using Runge-Kutta integration

Runge-Kutta integration

Runge-Kutta integration provides more accurate values for a given level of computational effort than Euler integration, for those models involving only systems of differential equations. Simile always uses Euler integration by default, because it is applicable to any model.

For example, consider the equation dy/dt = t, and given y = 0 at t = 0. In this case, there is an analytical solution, y = ½t2, and therefore at t = 100, y = 5000. Using the Euler integration algorithm, and using a step-size of 0.1 (the default) Simile calculates the answer to be 4995. Using Runge-Kutta integration, the correct answer of 5000 is found using the same step size (and therefore approximately the same computational effort). Even using a step size of 0.01 the answer is calculated to be 4999.5, using Euler integration.

However, it must be noted that Runge-Kutta integration is not always applicable, unlike Euler integration, which can be used to solve any system of differential or difference equations.

In particular, Simile allows the construction of difference equations, where certain quantities accumulate in sudden increments corresponding to weeks, months or years, for example. Interest payments, in particular, correspond to this method of accumulation. These systems must be solved by Euler integration, with the step size chosen to correspond to the periodicity.

Furthermore, although Runge-Kutta integration does not fail when used with discontinuous functions (such as are involved in the creation and destruction of instances of population submodels) the results will be no more precise than using Euler integration.

The choice between Euler and Runge-Kutta integration is made at run time, using a drop-down box in the Run Control. The chosen method is saved with the model and restored when the model is loaded. The user can experiment with running a model using either method, without having to rebuild the model, to understand whether a significant improvement is achieved.

In: Contents >> Running models