R/old/TODO.md

In mlysy/msde: Bayesian Inference for Multivariate Stochastic Differential Equations

All tasks have been done.

Keep track of TODO lists

1. Finish testthat test for sde.pf. See why there is a discrepancy, perhaps due to invalid proposals? See validx(x, theta) in e.g., test-hest_sd.R. Also, randomize inputs as much as possible (can do for other tests as well).

2. Make sde.pf usable. First of all, no need to return all X/lwgt output. Last observation is sufficient. Can potentially add a logical argument to output if desired. Second, don't always provide Z matrix. In fact, sdeFilter constructor can overload without Z, which requires significantly less memory (i.e., only enough for one obs), i.e., much faster if we rerun sde.pf over and over, realllocating memory every time. Also, provide R-level argument to smc::sampler RESAMPLE type (no resampling + prespecified Z is mainly for debugging, though does have other uses). See RESAMPLE argument in RcppSMC, as it explains how/when to resample.

3. Example in documentation.

4. augment this to sde.pf, i.e, any sdeModel.

   • make sdePF <= particleEval a virtual function of sdeCobj/sdeRobj
   • every call to sde.pf allocates/deallocates full memory.
   • sdeSMC.cpp should become (possible multiple) header files.

5. add smc debug to each model using msde-test_debug.R

6. testthat check deterministic Z input, history = TRUE/FALSE.

7. Test a PMCMC written in R.

   • You can check against a so-called multivariate Ornstein-Uhlenbeck model, for which Kalman filter analytically does filter calculation. This is coded in \code{mou.loglik}.
   • Compare speed/accuracy against \code{sde.post} using true posterior as benchmark.
   • Write up a report in the form of a vignette, which also shows how to use \code{sde.pf}.

8. finish debugging

9. Z should be passed as 3d array at R level, and data output should also be a 3d array.

mlysy/msde documentation built on May 28, 2022, 5:18 p.m.