pboesu / debinfer Goto Github PK

The PBSddesolve package is not available for download anymore, which is causing this dependency clash. Is there a way around it?
Thanks

Dear Dr. Boersch-Supan (and collaborators).

Quick question: is there a simple way to perform inference for lists of parameters?

For example, in a consumer resource model each species has it's own growth rate. In deSolve I can pass this in as a list and evaluate a series of equations using a loop.

My first intuition was to initialise multiple values when declaring the parameter:

mu <- debinfer_par(name = "mu", var.type = "de", fixed = FALSE,
          value = rep(0.5, n),
          prior = "norm", hypers = list(mean = 0, sd = 1), 
          prop.var = 0.0001, samp.type="rw")

My second is to write everything out long-hand mu1, mu2, ... which I'd like to avoid.

Any help appreciated,

Cheers,
Andrew

[1] "proposal outside prior support for Tmin = 1.79816012912371"
DLSODA-  ITASK = 4 or 5 and TCRIT (=R1) behind TCUR (=R2)   
In above message, R1 = 7, R2 = 7

Error in lsoda(y, times, func, parms, lags = control, ...) : 
  illegal input detected before taking any integration steps - see written message
 [1] "Solver failed with current state = 1.12203262641034 and proposal ds = 1.50219312234621"  "Solver failed with current state = 0.913153225090355 and proposal ds = 1.50219312234621"
 [3] "Solver failed with current state = 1.35363732916556 and proposal ds = 1.50219312234621"  "Solver failed with current state = 0.673927449394947 and proposal ds = 1.50219312234621"
 [5] "Solver failed with current state = 15.573702506406 and proposal ds = 1.50219312234621"   "Solver failed with current state = 2.42665556177606 and proposal ds = 1.50219312234621" 
 [7] "Solver failed with current state = 120 and proposal ds = 1.50219312234621"               "Solver failed with current state = 0 and proposal ds = 1.50219312234621"                
 [9] "Solver failed with current state = 0 and proposal ds = 1.50219312234621"                 "Solver failed with current state = -443.989025874076 and proposal ds = 1.50219312234621"
[1] "posterior not finite for proposal ds = 1.50219312234621"

The sample vector should be collapsed before pasting together the error message

https://github.com/pboesu/debinfer/blob/95ea267/R/de_mcmc_rev.R#L41
Prefer vapply(all.params, function(x) x$name, character(1)) to get better behaviour on unexpected input (e.g, all.params=NULL).

https://github.com/pboesu/debinfer/blob/95ea267/R/de_mcmc_rev.R#L57 (and elsewhere)
Don't use print() where you mean message(), and consider making conditional behind a verbose argument.

How to use truncated normal distribution as a prior in deBInfer?

Throughout the code examples in the paper/vignettes, be consistent with use of <- and =. (e.g., p.10, l. 185)

Some characters in the example code in the manuscript PDF are replaced by spaces when copying and pasting.

HPD is not plotted when the default value for color is overwritten by a single color name

Within R, I can find the vignettes in the doc directory but they don't seem accessible via the vignette command. For example, after loading the deBInfer package, the following command
vignette("logistic_ode_example",package="deBInfer")
results in the warning message...
vignette ‘logistic_ode_example’ not found

Does the package work with compiled ODE models (as described in vignette("compiledCode")? I didn't chase this throughm but it looks like it from the look of solve_de, but not from the definition of de_mcmc. This would probably be worth documenting and testing because by the time you're doing inference with an ODE/DDE system users would probably want the equations as fast as possible. If you do do this, what sort of speed up do you see?

see here: travis-ci/travis-ci#6139

The solver system is not open to extension in its current state:

  if(solver == 1 || solver == "ode"){
    #require(deSolve)
    out <- try(ode(inits, times, sim, parms=params, verbose=verbose , method=method, ...))
  }

Is there a reason why solver can't be made more generic? All it would really take is passing solver in as a function (along with a description of the form it should take). The existing ode/dde/dede if/else statements could be largely replaced by a little bit of wrapper code, such as:

if (is.character(solver)) {
  if (solver == "ode") {
    solver <- deSolve::ode(inits, times, sim, params, 
                           verbose=verbose, method=method, ...)
  } # ...and so on for the others.
}

In any case, because the method parameters passed along is the same as the default to ode, it can be dropped and passed along with ....

A suggestion to add a CoC.

Thank you so for much for this promising-looking package!

I was wondering how one would go about defining initial conditions that depend on each other. In population dynamics with a constant number of individuals, e.g.

dA/dt = -A
dB/dt = +A,

the total number of individuals is given by N = A(t) + B(t) = const at all times . If I want to use B(0) as a parameter to infer, is there any way to tell the package that at all times

A(0) = N - B(0)

while B(0) is not fixed?

Caution applies with 1:n (prefer seq_len(n)) and 1:length(x) (prefer seq_along(n)).

Please add in an auto.layout option.

https://github.com/pboesu/debinfer/blob/95ea267/R/de_mcmc_rev.R#L37
This seems overly restrictive; it would be nicer if you could use deSolve-style positional argument matching rather than require specific argument names.

this, and a straightforward command to simulate & plot from initial values would be helpful for troubleshooting ill-behaved DE models.

A suggested addition to the DDE vignette - to present all the posterior predictive trajectories on a single figure e.g.

plot(chytrid, xlab="Time (days)", ylab= "Zoospores x 10e4", xlim=c(0,10))

for(i in 1:100){
  DATA1 <- as.data.frame(post_traj$sims[i])
  lines(DATA1[,2] ~ DATA1[,1])
  lines(DATA1[,3] ~ DATA1[,1],col="red")
  lines(DATA1[,4] ~ DATA1[,1],col="blue")
}