mcreel / selectstatisticsabc.jl Goto Github PK

Instructions for replicating the results in  "On Selection of Statistics for Approximate Bayesian Computing (or the Method of Simulated Moments)", Michael Creel and Dennis Kristensen, 2015, Computational Statistics & Data Analysis, http://dx.doi.org/10.1016/j.csda.2015.05.005

Note: I recommend the methods presented in https://github.com/mcreel/NeuralNetsForIndirectInference.jl over the ones presented here. Those methods smoothly weight statistics, rather that selecting statistics. That option is simpler, faster, and has better performance, in my experience.

This is updated code,  which can use either local constant or local linear kernel regression. The default is local constant, for speed. To obtain the exact code used in the paper, checkout version 1.0 of  this repository.

files:
make_simdata.jl: used to make data files for the linear regression
                 example
Select.jl:       the selection program, serial version
Select_mpi.jl:   the selection program, parallel version
SelectionAlgorithm.jl: the main algorithm: the objective
                function, the simulated annealing algorithm, and
                the K nearest neighbors smoother for computing 
                the ABC estimator.

To create the data for the linear regression example by
typing, at the Julia prompt
     include("make_simdata.jl")

To perform selection of statistics:

The file Select.jl computes one replication of the
selection algorithm, using the simdata.30 data set.
Run this by typing the following line, at the Julia prompt
     include("Select.jl")

To perform selection in parallel, Select_mpi.jl
performs 100 replications of the method, again using
the simdata.30 data set. To run this, execute
    mpirun -np X julia Select_mpi.jl
from the system prompt. X should be less than the
total compute cores, and should be an even divisor
of 100.

Analyze.m is a GNU Octave script that analyzes the
results of multiple runs, picking out the best set of
statistics, and giving other information.
Use it via "octave --eval Analyze"

To repeat this for the sample size n=100, edit
make_simdata.jl appropriately, then repeat
the above instructions.

To perform selection for the jump-diffusion model,
please contact me for the data set.

For any doubts, please write [email protected]