BFGS Solver

Creates a solver using the BFGS quasi-Newton method via optim(). BFGS approximates the Hessian from gradient information, providing second-order-like convergence without computing the Hessian directly.

Usage

bfgs(max_iter = 100L, tol = 1e-08, report = 0L)

Arguments

max_iter

Maximum number of iterations

tol

Convergence tolerance (passed to optim's reltol)

report

Reporting frequency (0 = no reporting)

Value

A solver function with signature (problem, theta0, trace) -> mle_result

Details

BFGS is often a good default choice: it's more robust than Newton-Raphson (no matrix inversion issues) and faster than gradient ascent (uses curvature information).

The solver automatically uses the score function from the problem if available, otherwise computes gradients numerically.

Examples

if (FALSE) { # \dontrun{
# Basic usage
result <- bfgs()(problem, c(0, 1))

# Race BFGS against gradient ascent
strategy <- bfgs() %|% gradient_ascent()
} # }