HGF.jl

This repo holds a minimal implementation of the Hierarchical Gaussian Filter (HGF) for analysing behavioral data in julia.

Warning: This is untested and experimental research software, please do not use it unless you know what you are doing.

Installation

You can install the package with the following command (from the julia REPL):

        ]add https://github.com/toreerdmann/HGF.jl

Usage

using HGF
using Distributions: Normal, Exponential
using Plots

# draw design
ps = repeat([.2, .8, .2], inner = 100)
u  = [rand() < p for p in ps]

# Define a model
M = Model(BinaryeHGF(mu_0 = [NaN, 1, Exponential(1)],
                     sa_0 = [NaN, .1, .1],
                     om   = [NaN, Normal(-5, 1)],
                     th   = .1),
          SoftmaxBinary())

# Draw a model from the prior and run it forward
m = draw(M)
traj = m.pm(u)
plt = scatter(u .+ randn(length(u)) .* .01, legend=:none)
plot!(plt, traj.muhat[:,1]) 

# draw subjects
nsubs = 10
subs = [draw(M) for rep in 1:nsubs]
# simulate from each subject
ysim = [m(u) for m in subs]

# fit data
fits = [fitModel(M, u, ysim[i],
                 Options(niter = 200))[1] for i in 1:nsubs]

# look at simulated parameter vs. estimates 
omtrue = [sub.pm.om[2] for sub in subs]
omest  = [fit.pm.om[2].μ for fit in fits]
omest_err  = [fit.pm.om[2].σ for fit in fits]
scatter(omtrue, omest, yerror=omest_err)
plot!(x -> x)