News

I am happy to say that a new version of FLightR is now available on CRAN.

This release mainly focuses on speed. The particle filter in FLightR can be computationally demanding, especially when using many particles or large spatial grids, and this version makes those runs much faster.

The biggest change is a new cached and partially cached propagation backend. In simple terms, FLightR now avoids recalculating many of the same distances and bearings again and again during the particle filter. This makes a large difference. In one full-length validation run with 1 million particles and 627 twilight events, the old implementation took roughly 7-9 hours. The new version completed the same kind of run in about 24 minutes, while giving comparable likelihoods and posterior summaries.

The improvement is especially useful for larger analyses, where the old runtime could easily become a practical bottleneck. On a larger realistic spatial grid, the new partial cached backend matched the legacy backend in likelihood and checked posterior summaries, while running several times faster.

There are also some smaller but useful changes. The package is now quieter by default: run.particle.filter() uses verbose = FALSE, while verbose = TRUE or verbose = "debug" can be used when progress messages or detailed diagnostics are needed. We also found that the old PSOCK parallelisation inside the particle filter can actually make things slower, so for now threads = 1 is recommended for run.particle.filter(). This does not mean that all parallelisation is bad; the preparation steps before the particle filter can still benefit from multiple cores.

You can install the new version from CRAN:

install.packages("FLightR")

The source code and release notes are available on GitHub.