Re: [LibBi] Are inline expressions of less precision than states variables?

[Sebastian Funk]

Hi Sohrab,

Can you provide a reproducible example? The model you quote does not compile as it stands.

Seb.

Sohrab Salehi wrote:
> Hi,
>
> I'm facing what is most probably an overflow problem. To impose boundary
> conditions, in the transition block,
> I assign the updated value of the state variable to a temp variable,
> declared via inline.
> But this produces NaNs after some timestep (say t_s) in simulating from the
> model,
> where combining the result of the temp variable at t_s with any number via
> any mathematical operation
> creates NaNs.
> When the same boundary condition is implemented using additional dummy STATE
> variable, it works as expected.
> This workaround would have been okay if it didn't lead to a doubling of the
> states to be inferred.
> I've made sure that enable-single is set to false.
>
> The model looks as follows:
> model someModel {
> dim k(size = 2)
> dim i(size = 3)
> const h = 0.001 // step size
> const n = 200
> const epsilon = 0.05 // observation error tolerance
> const xSum0_1 = 0;
> noise deltaW[k] // wienr process steps
> param x0[i] // workaround for starting value of x
> state x[k]
> obs y[k]
>
> sub parameter {
> x0[i] ~ gamma()
> }
>
> sub initial {
> x[k] <- x0[k]/(x0[0] + x0[1] + x0[2])
> }
>
> sub transition(delta = h) {
> deltaW[k] ~ wienr()
>
> inline xSum0_1 = 0;
> inline x_new0_1 = x[0] + (-.2*x[0] - .3*x[1] + .2)*n*x[0]*h + sqrt((1 - x
> [0])*x[0])*deltaW[0];
> inline x_new0_2 = min(1-xSum0_1, min(1, max(x_new0_1, 0)));
> inline xSum1_1 = xSum0_1 + x_new0_2;
> inline x_new1_1 = x[1] + (-.2*x[0] - .3*x[1] + .3)*n*x[1]*h + sqrt((1 - x[0]
> - x[1])*x[1]/(1 - x[0]))*deltaW[1] -
> sqrt((1 - x[0])*x[0])*deltaW[0]*x[1]/(1 - x[0]);
> inline x_new1_2 = min(1-xSum1_1, min(1, max(x_new1_1, 0)));
> inline xSum2_1 = xSum1_1 + x_new1_2;
> x[0] <- x_nw0_2;
> x[1] <- x_nw1_2;
> }
>
> sub observation {
> y[k] ~ uniform(x[k]-epsilon, x[k]+epsilon)
> }
> }
>
> I was wondering if you can think about any solutions?
>
> I'm using LibBi 1.3.0 and running the model using RBi.
>
> Thanks,
> Sohrab