graDiEnt 1.1.0
New features
- Box constraints:
GetAlgoParams gains
lower and upper parameters for per-parameter
bounds. Enforced via bounds_type = "reflect" (default,
mirrors proposals back across boundary) or "clip"
(truncates to boundary).
- Warm start:
optim_SQGDE gains
warm_start parameter. Pass the output of a previous run to
seed the population and continue optimization without
re-initializing.
- Disk recovery:
GetAlgoParams gains
recovery_path (.rds file path) and
recovery_freq (save interval in iterations). Partial
results are written to disk periodically via saveRDS,
surviving interrupts and OOM crashes. Load with
readRDS(recovery_path).
- Progress message control:
GetAlgoParams gains trace_print_freq. Set to
an integer to print progress every N iterations, or Inf to
suppress all iteration messages.
- Unified adaptation schemes:
rand,
current, and best DE schemes are now selected
via adapt_scheme in GetAlgoParams rather than
separate internal functions.
crossover_rate = 0: Now accepted;
selects exactly one random parameter to update per iteration.jitter_size = 0: Now accepted;
disables jitter noise.- Block coordinate updating:
GetAlgoParams gains param_block_list, an
optional list of integer vectors partitioning 1:n_params
into blocks. The optimizer cycles through blocks in order across
iterations, updating only the current block’s parameters instead of
using random crossover selection. Useful for high-dimensional problems
where updating parameter subsets per iteration is more efficient.
Bug fixes
- Fixed division-by-zero in
grad_approx_fn when two
particles are identical or have equal weights (zero-diff guard).
- Fixed
rand scheme incorrectly sampling
2*n_diff parents instead of 2*n_diff + 1,
causing out-of-bounds index access.
- Fixed
runif jitter applied over n_params
elements instead of length(param_indices) when
crossover_rate < 1.
- Fixed
GetAlgoParams validation for
n_params, n_particles, and
n_iter: finiteness is now checked before
as.integer() coercion, preventing silent NA
production and spurious “condition has length > 1” errors on vector
inputs.
- Fixed early stopping crashing (and computing a wrong
'percent' metric) when thin > 1: the
convergence-check window is now clamped to the iterations stored so
far.
graDiEnt 1.0.1
- Package uses “message()” instead of “print()” to communicate with
user.
- minor fix to documentation.
graDiEnt 1.0.0
- First public release of package.