Automatic differentiation functionality using dual numbers has been
integrated with Optim with the "autodiff" keyword option and released
as version 0.2.0:
rosenbrock(x::Vector) = (1.0 - x[1])^2 + 100.0 * (x[2] - x[1]^2)^2
optimize(rosenbrock, [0.0, 0.0], method = :l_bfgs, autodiff = true)
When this option is set, exact numerical derivatives will be computed
by passing in a Vector{Dual{Float64}} to the input function. The
function must be written to accept a generic Vector, and should be
type stable (e.g. using zero(eltype(x)) instead of 0.0 for
accumulators). Using dual numbers has about the same cost as computing
gradients via finite differences, but the resulting gradient has much
greater numerical accuracy. It is still more efficient to provide your
own gradient function if available. Reverse-mode AD is in planning.
We encourage Optim users to test out this functionality and report any
issues; we're considering making it the default.
Miles
On Wed, Dec 11, 2013 at 5:23 PM, Miles Lubin <[email protected]> wrote:
> PR to make this work transparently with Optim:
> https://github.com/JuliaOpt/Optim.jl/pull/39
