function [traj, infStates] = tapas_hgf_whatworld(r, p, varargin)
% Calculates the trajectories of the agent's representations under the HGF
%
% This function can be called in two ways:
%
% (1) tapas_hgf_whatworld(r, p)
%
% where r is the structure generated by tapas_fitModel and p is the parameter vector in native space;
%
% (2) tapas_hgf_whatworld(r, ptrans, 'trans')
%
% where r is the structure generated by tapas_fitModel, ptrans is the parameter vector in
% transformed space, and 'trans' is a flag indicating this.
%
% --------------------------------------------------------------------------------------------------
% Copyright (C) 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/>.
% Check whether we have a configuration structure
if ~isfield(r,'c_prc')
error('tapas:hgf:ConfigRequired', 'Configuration required: before calling tapas_hgf_whatworld, tapas_hgf_whatworld_config has to be called.');
end
% Transform paramaters back to their native space if needed
if ~isempty(varargin) && strcmp(varargin{1},'trans');
p = tapas_hgf_whatworld_transp(r, p);
end
% Number of states whose contingencies have to be learned
ns = r.c_prc.n_states;
% Number of elements of the transition matrix
ntr = ns^2;
% Unpack parameters
mu2_0 = reshape(p(1:ntr)',ns,ns);
sa2_0 = reshape(p(ntr+1:2*ntr)',ns,ns);
mu3_0 = p(2*ntr+1);
sa3_0 = p(2*ntr+2);
ka = p(2*ntr+3);
om = p(2*ntr+4);
th = p(2*ntr+5);
% Add dummy "zeroth" trial
u = [1; r.u(:,1)];
% Number of trials (including prior)
n = length(u);
% Initialize updated quantities
% Representations
mu1 = NaN(n,ns,ns);
pi1 = NaN(n,ns,ns);
mu2 = NaN(n,ns,ns);
pi2 = NaN(n,ns,ns);
mu3 = NaN(n,1);
pi3 = NaN(n,1);
% Other quantities
mu1hat = NaN(n,ns,ns);
pi1hat = NaN(n,ns,ns);
mu2hat = NaN(n,ns,ns);
pi2hat = NaN(n,ns,ns);
mu3hat = NaN(n,1);
pi3hat = NaN(n,1);
v2 = NaN(n,1);
w2 = NaN(n,ns,ns);
da1 = NaN(n,ns,ns);
da2 = NaN(n,ns,ns);
% Representation priors
% Note: first entries of the other quantities remain
% NaN because they are undefined and are thrown away
% at the end; their presence simply leads to consistent
% trial indices.
mu1(1,:,:) = tapas_sgm(mu2_0, 1);
pi1(1,:,:) = 1./(mu1(1,:,:).*(1-mu1(1,:,:)));
mu2(1,:,:) = mu2_0;
pi2(1,:,:) = 1./sa2_0;
mu3(1) = mu3_0;
pi3(1) = 1/sa3_0;
% Pass through representation update loop
for k = 2:1:n
if not(ismember(k-1, r.ign))
%%%%%%%%%%%%%%%%%%%%%%
% Effect of input u(k)
%%%%%%%%%%%%%%%%%%%%%%
% Note: there is only an effect on the column of
% the transition matrix that corresponds to the
% previous outcome u(k-1)
% 1st level
% ~~~~~~~~~
% Unnormalized predictions
mu1hat(k,:,:) = squeeze(tapas_sgm(mu2(k-1,:,:), 1));
% Precisions of predictions
pi1hat(k,:,:) = 1./(mu1hat(k,:,:).*(1 -mu1hat(k,:,:)));
% Posterior for each possible transition
mu1(k,:,u(k-1)) = 0;
mu1(k,u(k),u(k-1)) = 1;
% Precision of posterior for each possible
% transition is infinite owing to absence
% of noise
pi1(k,:,u(k-1)) = Inf;
% Prediction errors
da1(k,:,u(k-1)) = mu1(k,:,u(k-1)) -mu1hat(k,:,u(k-1));
% 2nd level
% ~~~~~~~~~
% Predictions
mu2hat(k,:,:) = mu2(k-1,:,:);
% Precisions of predictions
pi2hat(k,:,:) = 1./(1./pi2(k-1,:,:) +exp(ka *mu3(k-1) +om));
% Updates
% Without observation, pi2 is equal to pi2hat
pi2(k,:,:) = pi2hat(k,:,:);
% However, where we have an observation, the usual update applies
pi2(k,:,u(k-1)) = pi2hat(k,:,u(k-1)) +1./pi1hat(k,:,u(k-1));
% By default, carry means of predictions forward
mu2(k,:,:) = mu2hat(k,:,:);
% However, where we have a prediction error, perform an update
mu2(k,:,u(k-1)) = mu2hat(k,:,u(k-1)) +1./pi2(k,:,u(k-1)) .*da1(k,:,u(k-1));
% Volatility prediction errors
da2(k,:,u(k-1)) = (1./pi2(k,:,u(k-1)) +(mu2(k,:,u(k-1)) -mu2hat(k,:,u(k-1))).^2) .*pi2hat(k,:,u(k-1)) -1;
% 3rd level
% ~~~~~~~~~
% Prediction
mu3hat(k) = mu3(k-1);
% Precision of prediction
pi3hat(k) = 1/(1/pi3(k-1) +th);
% Weighting factors
v2(k) = exp(ka *mu3(k-1) +om);
w2(k,:,u(k-1)) = v2(k) *pi2hat(k,:,u(k-1));
% Updates
pi3(k) = pi3hat(k) +sum(1/2 *ka^2 *w2(k,:,u(k-1)) .*(w2(k,:,u(k-1)) +(2 *w2(k,:,u(k-1)) -1) .*da2(k,:,u(k-1))));
if pi3(k) <= 0
error('tapas:hgf:NegPostPrec', 'Negative posterior precision. Parameters are in a region where model assumptions are violated.');
end
mu3(k) = mu3hat(k) +sum(1/2 *1/pi3(k) *ka *w2(k,:,u(k-1)) .*da2(k,:,u(k-1)));
else
mu1(k,:,:) = mu1(k-1,:,:);
pi1(k,:,:) = pi1(k-1,:,:);
mu2(k,:,:) = mu2(k-1,:,:);
pi2(k,:,:) = pi2(k-1,:,:);
mu3(k) = mu3(k-1);
pi3(k) = pi3(k-1);
mu1hat(k,:,:) = mu1hat(k-1,:,:);
pi1hat(k,:,:) = pi1hat(k-1,:,:);
mu2hat(k,:,:) = mu2hat(k-1,:,:);
pi2hat(k,:,:) = pi2hat(k-1,:,:);
mu3hat(k) = mu3hat(k-1);
pi3hat(k) = pi3hat(k-1);
v2(k) = v2(k-1);
w2(k,:,:) = w2(k-1,:,:);
da1(k,:,:) = da1(k-1,:,:);
da2(k,:,:) = da2(k-1,:,:);
end
end
% Remove representation priors
mu1(1,:,:) = [];
pi1(1,:,:) = [];
mu2(1,:,:) = [];
pi2(1,:,:) = [];
mu3(1) = [];
pi3(1) = [];
% Remove other dummy initial values
mu1hat(1,:,:) = [];
pi1hat(1,:,:) = [];
mu2hat(1,:,:) = [];
pi2hat(1,:,:) = [];
mu3hat(1) = [];
pi3hat(1) = [];
v2(1) = [];
w2(1,:,:) = [];
da1(1,:,:) = [];
da2(1,:,:) = [];
% Create result data structure
traj = struct;
traj.mu = NaN(n-1,3,ns,ns);
traj.mu(:,1,:,:) = mu1;
traj.mu(:,2,:,:) = mu2;
traj.mu(:,3,1,1) = mu3;
traj.sa = NaN(n-1,3,ns,ns);
traj.sa(:,1,:,:) = 1./pi1;
traj.sa(:,2,:,:) = 1./pi2;
traj.sa(:,3,1,1) = 1./pi3;
traj.muhat = NaN(n-1,3,ns,ns);
traj.muhat(:,1,:,:) = mu1hat;
traj.muhat(:,2,:,:) = mu2hat;
traj.muhat(:,3,1,1) = mu3hat;
traj.sahat = NaN(n-1,3,ns,ns);
traj.sahat(:,1,:,:) = 1./pi1hat;
traj.sahat(:,2,:,:) = 1./pi2hat;
traj.sahat(:,3,1,1) = 1./pi3hat;
traj.v = v2;
traj.w = w2;
traj.da = NaN(n-1,2,ns,ns);
traj.da(:,1,:,:) = da1;
traj.da(:,2,:,:) = da2;
% Updates with respect to prediction
traj.ud = traj.mu -traj.muhat;
% Psi (precision weights on prediction errors)
psi = NaN(n-1,3,ns,ns);
for k = 1:n-1
psi(k,2,:,:) = 1./pi2(k,:,:);
psi(k,3,:,:) = pi2hat(k,:,:)./pi3(k);
end
traj.psi = psi;
% Epsilons (precision-weighted prediction errors)
epsi = NaN(n-1,3,ns,ns);
for k = 1:n-1
epsi(k,2,:,:) = squeeze(psi(k,2,:,:)) .*squeeze(da1(k,:,:));
epsi(k,3,:,:) = squeeze(psi(k,3,:,:)) .*squeeze(da2(k,:,:));
end
traj.epsi = epsi;
% Implied learning rates at the first level
lr1 = NaN(n-1,ns,ns);
for k = 1:n-1
upd1 = tapas_sgm(mu2(k,:,:), 1) -mu1hat(k,:,:);
lr1(k,:,:) = upd1./da1(k,:,:);
end
% Full learning rate (full weights on prediction errors)
wt = NaN(n-1,3,ns,ns);
wt(:,1,:,:) = lr1;
wt(:,2,:,:) = psi(:,2,:,:);
v2psi = NaN(n-1,ns,ns);
for k = 1:n-1
v2psi(k,:,:) = v2(k)*psi(k,3,:,:);
end
wt(:,3,:,:) = 1/2 *ka *v2psi;
traj.wt = wt;
% Create matrices for use by the observation model
infStates = NaN(n-1,3,ns,ns,4);
infStates(:,:,:,:,1) = traj.muhat;
infStates(:,:,:,:,2) = traj.sahat;
infStates(:,:,:,:,3) = traj.mu;
infStates(:,:,:,:,4) = traj.sa;
return;