search

jump-dev / Ipopt.jl

A Julia interface to the Ipopt nonlinear solver
https://github.com/coin-or/ipopt
Other
152 stars 58 forks source link

Turning off AD #432

Closed ioannisPApapadopoulos closed 1 month ago

ioannisPApapadopoulos commented 1 month ago

Hi! I am trying to recreate some bad behavior I used to have with Ipopt when the Hessian was being approximated by BFGS rather than AD. Currently the AD Hessian works beautifully and I would like to contrast it.

Is there a way to turn off the AD and force Ipopt to use the BFGS default in this wrapper?

Thank you in advance!

frapac commented 1 month ago

You can switch to the LBFGS algorithm implemented in Ipopt using the option hessian_approximation. E.g. 

model = Model(Ipopt.Optimizer)
JuMP.set_optimizer_attribute(model, "hessian_approximation", "limited-memory")

The performance of LBFGS depends on the size of the history used in the approximation (by default, 6). Depending on your application you can increase that number quite significantly (usually the higher the better): 

JuMP.set_optimizer_attribute(model, "limited_memory_max_history", 10)

Please refer to Ipopt's documentation for more details.

ioannisPApapadopoulos commented 1 month ago

Perfect, thank you very much @frapac !