Hierarchical Gaussian Filter (HGF) model of conditioned hallucinations task (Powers et al 2017)

 Download zip file 
Help downloading and running models
Accession:229278
This is an instantiation of the Hierarchical Gaussian Filter (HGF) model for use with the Conditioned Hallucinations Task.
Reference:
1 . Powers AR, Mathys C, Corlett PR (2017) Pavlovian conditioning-induced hallucinations result from overweighting of perceptual priors. Science 357:596-600 [PubMed]
Citations  Citation Browser
Model Information (Click on a link to find other models with that property)
Model Type:
Brain Region(s)/Organism:
Cell Type(s):
Channel(s):
Gap Junctions:
Receptor(s):
Gene(s):
Transmitter(s):
Simulation Environment: MATLAB;
Model Concept(s): Hallucinations;
Implementer(s): Powers, Al [albert.powers at yale.edu]; Mathys, Chris H ;
/
HGF
analysis
hgfToolBox_condhalluc1.4
README
COPYING *
example_binary_input.txt
example_categorical_input.mat
example_usdchf.txt
Manual.pdf
tapas_autocorr.m
tapas_bayes_optimal.m
tapas_bayes_optimal_binary.m
tapas_bayes_optimal_binary_config.m
tapas_bayes_optimal_binary_transp.m
tapas_bayes_optimal_categorical.m
tapas_bayes_optimal_categorical_config.m
tapas_bayes_optimal_categorical_transp.m
tapas_bayes_optimal_config.m
tapas_bayes_optimal_transp.m
tapas_bayes_optimal_whatworld.m
tapas_bayes_optimal_whatworld_config.m
tapas_bayes_optimal_whatworld_transp.m
tapas_bayes_optimal_whichworld.m
tapas_bayes_optimal_whichworld_config.m
tapas_bayes_optimal_whichworld_transp.m
tapas_bayesian_parameter_average.m
tapas_beta_obs.m
tapas_beta_obs_config.m
tapas_beta_obs_namep.m
tapas_beta_obs_sim.m
tapas_beta_obs_transp.m
tapas_boltzmann.m
tapas_cdfgaussian_obs.m
tapas_cdfgaussian_obs_config.m
tapas_cdfgaussian_obs_transp.m
tapas_condhalluc_obs.m
tapas_condhalluc_obs_config.m
tapas_condhalluc_obs_namep.m
tapas_condhalluc_obs_sim.m
tapas_condhalluc_obs_transp.m
tapas_condhalluc_obs2.m
tapas_condhalluc_obs2_config.m
tapas_condhalluc_obs2_namep.m
tapas_condhalluc_obs2_sim.m
tapas_condhalluc_obs2_transp.m
tapas_Cov2Corr.m
tapas_datagen_categorical.m
tapas_fit_plotCorr.m
tapas_fit_plotResidualDiagnostics.m
tapas_fitModel.m
tapas_gaussian_obs.m
tapas_gaussian_obs_config.m
tapas_gaussian_obs_namep.m
tapas_gaussian_obs_sim.m
tapas_gaussian_obs_transp.m
tapas_hgf.m
tapas_hgf_ar1.m
tapas_hgf_ar1_binary.m
tapas_hgf_ar1_binary_config.m
tapas_hgf_ar1_binary_namep.m
tapas_hgf_ar1_binary_plotTraj.m
tapas_hgf_ar1_binary_transp.m
tapas_hgf_ar1_config.m
tapas_hgf_ar1_mab.m
tapas_hgf_ar1_mab_config.m
tapas_hgf_ar1_mab_plotTraj.m
tapas_hgf_ar1_mab_transp.m
tapas_hgf_ar1_namep.m
tapas_hgf_ar1_plotTraj.m
tapas_hgf_ar1_transp.m
tapas_hgf_binary.m
tapas_hgf_binary_condhalluc_plotTraj.m
tapas_hgf_binary_config.m
tapas_hgf_binary_config_startpoints.m
tapas_hgf_binary_mab.m
tapas_hgf_binary_mab_config.m
tapas_hgf_binary_mab_plotTraj.m
tapas_hgf_binary_mab_transp.m
tapas_hgf_binary_namep.m
tapas_hgf_binary_plotTraj.m
tapas_hgf_binary_pu.m
tapas_hgf_binary_pu_config.m
tapas_hgf_binary_pu_namep.m
tapas_hgf_binary_pu_tbt.m
tapas_hgf_binary_pu_tbt_config.m
tapas_hgf_binary_pu_tbt_namep.m
tapas_hgf_binary_pu_tbt_transp.m
tapas_hgf_binary_pu_transp.m
tapas_hgf_binary_transp.m
tapas_hgf_categorical.m
tapas_hgf_categorical_config.m
tapas_hgf_categorical_namep.m
tapas_hgf_categorical_norm.m
tapas_hgf_categorical_norm_config.m
tapas_hgf_categorical_norm_transp.m
tapas_hgf_categorical_plotTraj.m
tapas_hgf_categorical_transp.m
tapas_hgf_config.m
tapas_hgf_demo.m
tapas_hgf_demo_commands.m
tapas_hgf_jget.m
tapas_hgf_jget_config.m
tapas_hgf_jget_plotTraj.m
tapas_hgf_jget_transp.m
tapas_hgf_namep.m
tapas_hgf_plotTraj.m
tapas_hgf_transp.m
tapas_hgf_whatworld.m
tapas_hgf_whatworld_config.m
tapas_hgf_whatworld_namep.m
tapas_hgf_whatworld_plotTraj.m
tapas_hgf_whatworld_transp.m
tapas_hgf_whichworld.m
tapas_hgf_whichworld_config.m
tapas_hgf_whichworld_namep.m
tapas_hgf_whichworld_plotTraj.m
tapas_hgf_whichworld_transp.m
tapas_hhmm.m
tapas_hhmm_binary_displayResults.m
tapas_hhmm_config.m
tapas_hhmm_transp.m
tapas_hmm.m
tapas_hmm_binary_displayResults.m
tapas_hmm_config.m
tapas_hmm_transp.m
tapas_kf.m
tapas_kf_config.m
tapas_kf_namep.m
tapas_kf_plotTraj.m
tapas_kf_transp.m
tapas_logit.m
tapas_logrt_linear_binary.m
tapas_logrt_linear_binary_config.m
tapas_logrt_linear_binary_minimal.m
tapas_logrt_linear_binary_minimal_config.m
tapas_logrt_linear_binary_minimal_transp.m
tapas_logrt_linear_binary_namep.m
tapas_logrt_linear_binary_sim.m
tapas_logrt_linear_binary_transp.m
tapas_logrt_linear_whatworld.m
tapas_logrt_linear_whatworld_config.m
tapas_logrt_linear_whatworld_transp.m
tapas_ph_binary.m
tapas_ph_binary_config.m
tapas_ph_binary_namep.m
tapas_ph_binary_plotTraj.m
tapas_ph_binary_transp.m
tapas_quasinewton_optim.m
tapas_quasinewton_optim_config.m
tapas_riddersdiff.m
tapas_riddersdiff2.m
tapas_riddersdiffcross.m
tapas_riddersgradient.m
tapas_riddershessian.m
tapas_rs_belief.m
tapas_rs_belief_config.m
tapas_rs_precision.m
tapas_rs_precision_config.m
tapas_rs_precision_whatworld.m
tapas_rs_precision_whatworld_config.m
tapas_rs_surprise.m
tapas_rs_surprise_config.m
tapas_rs_transp.m
tapas_rs_whatworld_transp.m
tapas_rw_binary.m
tapas_rw_binary_config.m
tapas_rw_binary_dual.m
tapas_rw_binary_dual_config.m
tapas_rw_binary_dual_plotTraj.m
tapas_rw_binary_dual_transp.m
tapas_rw_binary_namep.m
tapas_rw_binary_plotTraj.m
tapas_rw_binary_transp.m
tapas_sgm.m
tapas_simModel.m
tapas_softmax.m
tapas_softmax_2beta.m
tapas_softmax_2beta_config.m
tapas_softmax_2beta_transp.m
tapas_softmax_binary.m
tapas_softmax_binary_config.m
tapas_softmax_binary_namep.m
tapas_softmax_binary_sim.m
tapas_softmax_binary_transp.m
tapas_softmax_config.m
tapas_softmax_namep.m
tapas_softmax_sim.m
tapas_softmax_transp.m
tapas_squared_pe.m
tapas_squared_pe_config.m
tapas_squared_pe_transp.m
tapas_sutton_k1_binary.m
tapas_sutton_k1_binary_config.m
tapas_sutton_k1_binary_plotTraj.m
tapas_sutton_k1_binary_transp.m
tapas_unitsq_sgm.m
tapas_unitsq_sgm_config.m
tapas_unitsq_sgm_mu3.m
tapas_unitsq_sgm_mu3_config.m
tapas_unitsq_sgm_mu3_transp.m
tapas_unitsq_sgm_namep.m
tapas_unitsq_sgm_sim.m
tapas_unitsq_sgm_transp.m
                            
function c = tapas_hgf_ar1_mab_config
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Contains the configuration for the Hierarchical Gaussian Filter (HGF) for AR(1)
% processes in multi-armed bandit situations. The template for such
% a situation is the task from
%
% Daw ND, O’Doherty JP, Dayan P, Seymour B, and Dolan RJ. (2006).
% Cortical substrates for exploratory decisions in humans. Nature, 441(7095), 876–879.
%
% The HGF is the model introduced in 
%
% Mathys C, Daunizeau J, Friston KJ, and Stephan KE. (2011). A Bayesian foundation
% for individual learning under uncertainty. Frontiers in Human Neuroscience, 5:39.
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% The recommended syntax for this model is
%
% >> est = tapas_fitModel(y, u, 'tapas_hgf_ar1_mabt_config', 'tapas_softmax_config');
%
% y here is the subject's choice (i.e., the number of the bandit chosen), u is an n-by-2 matrix
% where n is the number of trials. The first column is the payout on that trial, and the second
% column is again y. This has to appear in two places because the choice is relevant to the
% perceptual model.
%
% The HGF configuration consists of the priors of parameters and initial values. All priors are
% Gaussian in the space where the quantity they refer to is estimated. They are specified by their
% sufficient statistics: mean and variance (NOT standard deviation).
% 
% Quantities are estimated in their native space if they are unbounded (e.g., the omegas). They are
% estimated in log-space if they have a natural lower bound at zero (e.g., the sigmas).
% 
% The phis are estimated in 'logit space' because they are confined to the interval from 0 to 1.
% 'Logit-space' is a logistic sigmoid transformation of native space with a variable upper bound
% a>0:
% 
% tapas_logit(x) = ln(x/(a-x)); x = a/(1+exp(-tapas_logit(x)))
%
% Parameters can be fixed (i.e., set to a fixed value) by setting the variance of their prior to
% zero. Aside from being useful for model comparison, the need for this arises whenever the scale
% and origin at the j-th level are arbitrary. This is the case if the observation model does not
% contain the representations mu_j and sigma_j. A choice of scale and origin is then implied by
% fixing the initial value mu_j_0 of mu_j and either kappa_j-1 or omega_j-1.
%
% Fitted trajectories can be plotted by using the command
%
% >> tapas_hgf_ar1_plotTraj(est)
% 
% where est is the stucture returned by tapas_fitModel. This structure contains the estimated
% perceptual parameters in est.p_prc and the estimated trajectories of the agent's
% representations (cf. Mathys et al., 2011). Their meanings are:
%              
%         est.p_prc.mu_0       row vector of initial values of mu (in ascending order of levels)
%         est.p_prc.sa_0       row vector of initial values of sigma (in ascending order of levels)
%         est.p_prc.phi        row vector of phis
%         est.p_prc.m          row vector of ms
%         est.p_prc.ka         row vector of kappas (in ascending order of levels)
%         est.p_prc.om         row vector of omegas (in ascending order of levels)
%         est.p_prc.al         alpha
%
%         est.traj.mu          mu (rows: trials, columns: levels)
%         est.traj.sa          sigma (rows: trials, columns: levels)
%         est.traj.muhat       prediction of mu (rows: trials, columns: levels)
%         est.traj.sahat       precisions of predictions (rows: trials, columns: levels)
%         est.traj.v           inferred variance of random walk (rows: trials, columns: levels)
%         est.traj.w           weighting factors (rows: trials, columns: levels)
%         est.traj.da          volatility prediction errors  (rows: trials, columns: levels)
%         est.traj.dau         input prediction error
%         est.traj.ud          updates with respect to prediction  (rows: trials, columns: levels)
%         est.traj.psi         precision weights on prediction errors  (rows: trials, columns: levels)
%         est.traj.epsi        precision-weighted prediction errors  (rows: trials, columns: levels)
%         est.traj.wt          full weights on prediction errors (at the first level,
%                                  this is the learning rate) (rows: trials, columns: levels)
%
% Tips:
% - When analyzing a new dataset, take your inputs u and use
%
%   >> est = tapas_fitModel([], u, 'tapas_hgf_ar1_config', 'tapas_bayes_optimal_config');
%
%   to determine the Bayes optimal perceptual parameters (given your current priors as defined in
%   this file here, so choose them wide and loose to let the inputs influence the result). You can
%   then use the optimal parameters as your new prior means for the perceptual parameters.
%
% - If you get an error saying that the prior means are in a region where model assumptions are
%   violated, lower the prior means of the omegas, starting with the highest level and proceeding
%   downwards.
%
% - Alternatives are lowering the prior means of the kappas, if they are not fixed, or adjusting
%   the values of the kappas or omegas, if any of them are fixed.
%
% - If the log-model evidence cannot be calculated because the Hessian poses problems, look at
%   est.optim.H and fix the parameters that lead to NaNs.
%
% - Your guide to all these adjustments is the log-model evidence (LME). Whenever the LME increases
%   by at least 3 across datasets, the adjustment was a good idea and can be justified by just this:
%   the LME increased, so you had a better model.
%
% --------------------------------------------------------------------------------------------------
% Copyright (C) 2012-2013 Christoph Mathys, TNU, UZH & ETHZ
%
% This file is part of the HGF toolbox, which is released under the terms of the GNU General Public
% Licence (GPL), version 3. You can redistribute it and/or modify it under the terms of the GPL
% (either version 3 or, at your option, any later version). For further details, see the file
% COPYING or <http://www.gnu.org/licenses/>.


% Config structure
c = struct;

% Model name
c.model = 'hgf_ar1_mab';

% Number of bandits
c.n_bandits = 3;

% Number of levels (minimum: 2)
c.n_levels = 2;

% Input intervals
% If input intervals are irregular, the last column of the input
% matrix u has to contain the interval between inputs k-1 and k
% in the k-th row, and this flag has to be set to true
c.irregular_intervals = false;

% Sufficient statistics of Gaussian parameter priors

% PLACEHOLDER VALUES
% It is often convenient to set some priors to values
% derived from the inputs. This can be achieved by
% using placeholder values. The available placeholders
% are:
%
% 99991   Value of the first input
%         Usually a good choice for mu_0mu(1)
% 99992   Variance of the first 20 inputs
%         Usually a good choice for mu_0sa(1)
% 99993   Log-variance of the first 20 inputs
%         Usually a good choice for logsa_0mu(1)
%         and logalmu
% 99994   Log-variance of the first 20 inputs minus two
%         Usually a good choice for ommu(1)

% Initial mus and sigmas
% Format: row vectors of length n_levels
% For all but the first level, this is usually best
% kept fixed to 1 (determines origin on x_i-scale).
c.mu_0mu = [50, 1];
c.mu_0sa = [0, 0];

c.logsa_0mu = [log(70), log(0.1)];
c.logsa_0sa = [      0,        0];

% Phis
% Format: row vector of length n_levels.
% Phi is estimated in logit-space because it is
% bounded between 0 and 1
% Fix this to zero (leading to a Gaussian random walk) by
% setting logitphimu = -Inf; logitphisa = 0;
c.logitphimu = [tapas_logit(0.02,1), -Inf];
c.logitphisa = [            1,    0];

% ms
% Format: row vector of length n_levels.
% This should be fixed for all levels where the omega of
% the next lowest level is not fixed because that offers
% an alternative parametrization of the same model.
c.mmu = [ 50, c.mu_0mu(2)];
c.msa = [8^2,           0];

% Kappas
% Format: row vector of length n_levels-1.
% This should be fixed (preferably to 1) if the observation model
% does not use mu_i+1 (kappa then determines the scaling of x_i+1).
c.logkamu = [log(1)];
c.logkasa = [     0];

% Omegas
% Format: row vector of length n_levels
c.ommu = [  4,   -4];
c.omsa = [4^2,  4^2];

% Alpha
% Format: scalar
% Fix this to zero (no percpeptual uncertainty) by setting
% logalmu = -Inf; logalsa = 0;
c.logalmu = log(128);
c.logalsa = 0;

% Gather prior settings in vectors
c.priormus = [
    c.mu_0mu,...
    c.logsa_0mu,...
    c.logitphimu,...
    c.mmu,...
    c.logkamu,...
    c.ommu,...
    c.logalmu,...
         ];

c.priorsas = [
    c.mu_0sa,...
    c.logsa_0sa,...
    c.logitphisa,...
    c.msa,...
    c.logkasa,...
    c.omsa,...
    c.logalsa,...
         ];

% Check whether we have the right number of priors
expectedLength = 4*c.n_levels+2*(c.n_levels-1)+2;
if length([c.priormus, c.priorsas]) ~= 2*expectedLength;
    error('tapas:hgf:PriorDefNotMatchingLevels', 'Prior definition does not match number of levels.')
end

% Model function handle
c.prc_fun = @tapas_hgf_ar1_mab;

% Handle to function that transforms perceptual parameters to their native space
% from the space they are estimated in
c.transp_prc_fun = @tapas_hgf_ar1_mab_transp;

return;