Control numerical fitting in `gwrs`
Usage
gwrs_control(
tolerance = 1e-07,
max_iterations = 1000L,
n_threads = -1L,
grain_size = 16L,
keep_data = TRUE,
diagnostics = TRUE,
nonconvex_solver = c("coordinate", "guarded_block")
)Arguments
- tolerance
Relative coefficient convergence tolerance.
- max_iterations
Maximum number of solver iterations per location.
- n_threads
Number of parallel worker threads. Use `1` for serial execution or `-1` to let `RcppParallel` choose.
- grain_size
Minimum number of target locations assigned per parallel task.
- keep_data
Whether fitted objects retain the training matrices.
- diagnostics
Statistical diagnostic level. `"standard"` computes compact fit and effective-degrees-of-freedom diagnostics, `"full"` also retains quantities needed for local coefficient inference, and `"none"` skips statistical diagnostics. Logical values remain supported for backward compatibility (`TRUE` maps to `"standard"`). Numerical solver diagnostics are always retained because they are small and are required to audit convergence.
- nonconvex_solver
Gaussian SCAD/MCP solver: `"coordinate"` retains the legacy updates and failure behavior (default); `"guarded_block"` adds objective-decreasing pair and flat-tail QR proposals. In guarded mode a failed target stops its remaining lambda path. CV scores require convergence for every held-out observation. Other estimators ignore this option. No environment variable is needed.