yada_tailored_annealing: Do a tailored optimization using simulated annealing

In MichaelHoltonPrice/yada: Yet Another Demographic Analysis package

Description

A tailored optimization using simulated annealing that works well for univariate models. The annealing can be used by itself for optimization or to initialize another optimization algorithm. While this optimization via annealing usually requires more computations of the negative log-likelihood function compared to other algorithms (e.g., directly calling stats::optim with method='BFGS'), experience has shown that other algorithms do not always find the minimum of the negative log-likelihood. Put slightly differently, this tailored algorithm trades off number of computations for improved robustness, which is likely almost always a trade-off the user desires. The simulated annealing is adaptive. In particular, with each change in temperature (and prior to annealing at the first temperature) sampling is separately done to set the standard deviation of the (normal) proposal distribution to achieve an acceptance ratio of about 0.23.

Usage

yada_tailored_annealing(neg_log_lik, theta0, ...)

Arguments

neg_log_lik	The function for the negative log-likelihood
theta0	The initial value of the parameter vector
...	Variables required by neg_log_lik

Value

A list containing eta_best (the best value of the negative log-likelihood encountered), theta_best (the corresponding best parameter vector), and eta_vect (the full vector of negative log-likelihoods for the sampled parameter vectors)

MichaelHoltonPrice/yada documentation built on Sept. 19, 2021, 11:27 p.m.