search

traitecoevo / hmde

Implementation of hierarchical Bayesian longitudinal models to estimate differential equation parameters.
Other
2 stars 3 forks source link

vB model run failures #21

Closed Tess-LaCoil closed 3 months ago

Tess-LaCoil commented 6 months ago

https://github.com/traitecoevo/rmot/actions/runs/8785643601/job/24106637311#step:6:451

═ Failed tests ════════════════════════════════════════════════════════════════
── Failure ('test-rmot_models_vb.R:47:3'): Execution: vb single individual ─────
`par_ests` (`actual`) not equal to as.numeric(data$single_true_data$DE_pars[c(1, 2)]) (`expected`).

  `actual`: 2.7 7.2
`expected`: 0.1 7.2

Problem parameter is the growth coefficient. Doesn't always fail to estimate properly but every so often will.

Tess-LaCoil commented 6 months ago

https://github.com/traitecoevo/rmot/actions/runs/8778120858/job/24084079159 Succeeded here then failed later when attempting to merge in the case study datasets.

Tess-LaCoil commented 6 months ago

I am struggling to recreate this locally as it's not clear what the failure conditions are.

Tess-LaCoil commented 3 months ago

Through trial and error I've discovered that the MCMC algorithm manages to identify parameter combinations that are pathological for the RK4 algorithm at a given step size. The fix is to reduce the step size when running the individual model test as this will eliminate the pathological values. There's more work to be done here investigating the interaction because it's extremely reproducable, but that's not for this package.