Returns EDI's built-in policy table, consulted by the internal (non-exported)
dispatcher edi_cold_start_dispatch_policy(), for the smart_cold_start
default used by each inference class's C++ model-fitting backend. A TRUE
entry means the solver initializes via an OLS (or otherwise model-appropriate
heuristic) warm-up before iterating; FALSE means a plain zero-vector cold
start. Benchmarks show the OLS warm-up is net-negative for logistic and Poisson
IRLS at typical trial sizes (the one extra OLS solve costs more than the IRLS
iterations it saves), so those families — along with several G-computation-based
incidence/proportion inference classes — default to FALSE here.
Value
A named list with default (logical, the fallback when no
override pattern matches; TRUE in the built-in policy) and
inference_class_overrides (a named logical vector: regular-expression
pattern names to TRUE/FALSE values, matched against the
inference class name).
These TRUE/FALSE defaults are empirical performance
judgments computed on the maintainer's machine, not correctness facts —
the same heuristic can be net-positive or net-negative depending on
your hardware's core count, cache sizes, and BLAS backend. Run
tune_EDI_for_this_machine to re-measure this axis on your
own machine and persist any better setting it finds.
Details
The dispatcher checks the inference class name against
inference_class_overrides's named regular-expression patterns in list
order, returning the associated logical value at the first match; if none
match, it falls back to default (TRUE). Unlike
get_bootstrap_dispatch_policy, there is no separate
design-class-scoped override table here — only a single flat pattern list.
See also
get_bootstrap_dispatch_policy and
get_optimization_dispatch_policy for the analogous policies
controlling bootstrap CI type and default optimizer algorithm;
set_cold_start_dispatch_policy to override this policy at
runtime; tune_EDI_for_this_machine to re-benchmark it on
your own hardware.
Examples
get_cold_start_dispatch_policy()
#> $default
#> [1] TRUE
#>
#> $inference_class_overrides
#> ^InferenceIncidLogRegr$ ^InferencePropFractionalLogit$
#> FALSE FALSE
#> ^InferencePropGComp ^InferenceIncidGComp
#> FALSE FALSE
#> ^InferenceIncidKKGComp ^InferenceCountPoisson$
#> FALSE FALSE
#> ^InferenceCountQuasiPoisson$ ^InferenceCountRobustPoisson$
#> FALSE FALSE
#> ^InferenceIncidModifiedPoisson$
#> FALSE
#>