v1.0.2 - CRAN Submission

CRAN Submission Release

Changes since v1.0.1

Bug Fixes:

• Fix race() parallel execution with future package - worker processes now correctly load compositional.mle

Testing:

• Add unit tests for race(parallel = TRUE) functionality

Build:

• Add docs/ to .Rbuildignore to exclude pkgdown site from package tarball

R CMD check

0 errors ✔ | 0 warnings ✔ | 0 notes ✔

v1.0.1 - CRAN Submission

CRAN Submission Release

Changes since v1.0.0

Code Quality Improvements:

• Remove unused mle_config* functions (dead code cleanup)
• Fix CRAN policy: wrap message() in interactive() checks
• Add @return tags to all print methods
• Add @examples to utility functions
• Add @seealso cross-references to solvers

Bug Fixes:

• Fix input validation for mle_trace(every) parameter
• Fix unsafe float comparison in backtracking line search
• Add solver names to error messages for better debugging

Style:

• Fix line length violations (>80 chars)
• Fix indentation issues

R CMD check

0 errors ✔ | 0 warnings ✔ | 1 note (optional 'future' package)

compositional.mle v0.2.0

Major Refactoring: Composable MLE Solvers

This release renames the package from numerical.mle to compositional.mle and introduces a new composable API based on SICP principles.

Key Changes

• Factory Pattern: Solvers are now factory functions that return solver functions with uniform signature: (problem, theta0, trace) -> result
• Composition Operators:
  • %>>% - Sequential chaining (coarse-to-fine strategies)
  • %|% - Parallel racing (pick best result)
  • with_restarts() - Multiple random starting points
  • unless_converged() - Conditional refinement
• Problem Specification: New mle_problem() separates the statistical problem from optimization strategy
• New Solvers: bfgs(), lbfgsb(), nelder_mead(), random_search(), fisher_scoring()
• Tracing: mle_trace() for configurable iteration diagnostics
• Transformers: with_subsampling(), with_penalty() (L1/L2/elastic net)

Example

# Define problem once
problem <- mle_problem(loglike, score, constraint = mle_constraint(...))

# Create composable strategy
strategy <- grid_search(n = 5) %>>% gradient_ascent() %>>% newton_raphson()

# Or race different methods
strategy <- gradient_ascent() %|% bfgs() %|% nelder_mead()

# Solve
result <- strategy(problem, theta0)

Testing

All 255 tests passing.

v0.1.0 - Initial Release

numerical.mle v0.1.0

Initial release of the numerical.mle R package for numerical maximum likelihood estimation.

Features

Configuration System

• mle_config() - Base configuration for convergence criteria
• mle_config_gradient() - Gradient descent with fixed learning rate
• mle_config_linesearch() - Adaptive step size via backtracking line search
• mle_constraint() - Domain constraints with support checking and projection

Core Solvers

• mle_gradient_ascent() - First-order gradient-based optimization
• mle_newton_raphson() - Second-order optimization using Fisher information

Meta-Solvers

• mle_grid_search() - Exhaustive grid search over parameter space
• mle_random_restart() - Multiple random initializations for global optimization

Function Transformers

• with_subsampling() - Stochastic gradient ascent via mini-batching
• with_penalty() - Regularization with L1, L2, or elastic net penalties

Convenience Wrappers

• mle_grad() - Quick gradient ascent with sensible defaults
• mle_nr() - Quick Newton-Raphson with sensible defaults
• with_constraint() - Apply constraints to any solver

Installation

# Install from GitHub
devtools::install_github("queelius/numerical.mle")

Documentation

Full documentation available at: https://queelius.github.io/numerical.mle/