function [traj, infStates] = tapas_hgf_binary_mab(r, p, varargin)
% Calculates the trajectories of the agent's representations under the HGF in a multi-armed bandit
% situation with binary outcomes
%
% This function can be called in two ways:
%
% (1) tapas_hgf_binary_mab(r, p)
%
% where r is the structure generated by tapas_fitModel and p is the parameter vector in native space;
%
% (2) tapas_hgf_binary_mab(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/>.
% Transform paramaters back to their native space if needed
if ~isempty(varargin) && strcmp(varargin{1},'trans');
p = tapas_hgf_binary_mab_transp(r, p);
end
% Number of levels
try
l = r.c_prc.n_levels;
catch
l = (length(p)+1)/5;
if l ~= floor(l)
error('tapas:hgf:UndetNumLevels', 'Cannot determine number of levels');
end
end
% Number of bandits
try
b = r.c_prc.n_bandits;
catch
error('tapas:hgf:NumOfBanditsConfig', 'Number of bandits has to be configured in r.c_prc.n_bandits.');
end
% Coupled updating
% This is only allowed if there are 2 bandits. We here assume that the mu1hat for the two bandits
% add to unity.
coupled = false;
if r.c_prc.coupled == true
if b == 2
coupled = true;
else
error('tapas:hgf:HgfBinaryMab:CoupledOnlyForTwo', 'Coupled updating can only be configured for 2 bandits.');
end
end
% Unpack parameters
mu_0 = p(1:l);
sa_0 = p(l+1:2*l);
rho = p(2*l+1:3*l);
ka = p(3*l+1:4*l-1);
om = p(4*l:5*l-2);
th = exp(p(5*l-1));
% Add dummy "zeroth" trial
u = [0; r.u(:,1)];
y = [1; r.y(:,1)];
% Number of trials (including prior)
n = size(u,1);
% Construct time axis
if r.c_prc.irregular_intervals
if size(u,2) > 1
t = [0; r.u(:,end)];
else
error('tapas:hgf:InputSingleColumn', 'Input matrix must contain more than one column if irregular_intervals is set to true.');
end
else
t = ones(n,1);
end
% Initialize updated quantities
% Representations
mu = NaN(n,l,b);
pi = NaN(n,l,b);
% Other quantities
muhat = NaN(n,l,b);
pihat = NaN(n,l,b);
v = NaN(n,l);
w = NaN(n,l-1);
da = NaN(n,l);
% 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.
mu(1,1,:) = tapas_sgm(mu_0(2), 1);
pi(1,1,:) = Inf;
mu(1,2:end,:) = repmat(mu_0(2:end),[1 1 b]);
pi(1,2:end,:) = repmat(1./sa_0(2:end),[1 1 b]);
% Pass through representation update loop
for k = 2:1:n
if not(ismember(k-1, r.ign))
%%%%%%%%%%%%%%%%%%%%%%
% Effect of input u(k)
%%%%%%%%%%%%%%%%%%%%%%
% 2nd level prediction
muhat(k,2,:) = mu(k-1,2,:) +t(k) *rho(2);
% 1st level
% ~~~~~~~~~
% Prediction
muhat(k,1,:) = tapas_sgm(muhat(k,2,:), 1);
% Precision of prediction
pihat(k,1,:) = 1/(muhat(k,1,:).*(1 -muhat(k,1,:)));
% Updates
pi(k,1,:) = pihat(k,1,:);
pi(k,1,y(k)) = Inf;
mu(k,1,:) = muhat(k,1,:);
mu(k,1,y(k)) = u(k);
% Prediction error
da(k,1) = mu(k,1,y(k)) -muhat(k,1,y(k));
% 2nd level
% ~~~~~~~~~
% Prediction: see above
% Precision of prediction
pihat(k,2,:) = 1/(1/pi(k-1,2,:) +exp(ka(2) *mu(k-1,3,:) +om(2)));
% Updates
pi(k,2,:) = pihat(k,2,:);
pi(k,2,y(k)) = pihat(k,2,y(k)) +1/pihat(k,1,y(k));
mu(k,2,:) = muhat(k,2,:);
mu(k,2,y(k)) = muhat(k,2,y(k)) +1/pi(k,2,y(k)) *da(k,1);
% Volatility prediction error
da(k,2) = (1/pi(k,2,y(k)) +(mu(k,2,y(k)) -muhat(k,2,y(k)))^2) *pihat(k,2,y(k)) -1;
if l > 3
% Pass through higher levels
% ~~~~~~~~~~~~~~~~~~~~~~~~~~
for j = 3:l-1
% Prediction
muhat(k,j,:) = mu(k-1,j,:) +t(k) *rho(j);
% Precision of prediction
pihat(k,j,:) = 1/(1/pi(k-1,j,:) +t(k) *exp(ka(j) *mu(k-1,j+1,:) +om(j)));
% Weighting factor
v(k,j-1) = t(k) *exp(ka(j-1) *mu(k-1,j,y(k)) +om(j-1));
w(k,j-1) = v(k,j-1) *pihat(k,j-1,y(k));
% Updates
pi(k,j,:) = pihat(k,j,:) +1/2 *ka(j-1)^2 *w(k,j-1) *(w(k,j-1) +(2 *w(k,j-1) -1) *da(k,j-1));
if pi(k,j,1) <= 0
error('tapas:hgf:NegPostPrec', 'Negative posterior precision. Parameters are in a region where model assumptions are violated.');
end
mu(k,j,:) = muhat(k,j,:) +1/2 *1/pi(k,j) *ka(j-1) *w(k,j-1) *da(k,j-1);
% Volatility prediction error
da(k,j) = (1/pi(k,j,y(k)) +(mu(k,j,y(k)) -muhat(k,j,y(k)))^2) *pihat(k,j,y(k)) -1;
end
end
% Last level
% ~~~~~~~~~~
% Prediction
muhat(k,l,:) = mu(k-1,l,:) +t(k) *rho(l);
% Precision of prediction
pihat(k,l,:) = 1/(1/pi(k-1,l,:) +t(k) *th);
% Weighting factor
v(k,l) = t(k) *th;
v(k,l-1) = t(k) *exp(ka(l-1) *mu(k-1,l,y(k)) +om(l-1));
w(k,l-1) = v(k,l-1) *pihat(k,l-1,y(k));
% Updates
pi(k,l,:) = pihat(k,l,:) +1/2 *ka(l-1)^2 *w(k,l-1) *(w(k,l-1) +(2 *w(k,l-1) -1) *da(k,l-1));
if pi(k,l,1) <= 0
error('tapas:hgf:NegPostPrec', 'Negative posterior precision. Parameters are in a region where model assumptions are violated.');
end
mu(k,l,:) = muhat(k,l,:) +1/2 *1/pi(k,l,:) *ka(l-1) *w(k,l-1) *da(k,l-1);
% Volatility prediction error
da(k,l) = (1/pi(k,l,y(k)) +(mu(k,l,y(k)) -muhat(k,l,y(k)))^2) *pihat(k,l,y(k)) -1;
if coupled == true
if y(k) == 1
mu(k,1,2) = 1 -mu(k,1,1);
mu(k,2,2) = tapas_logit(1 -tapas_sgm(mu(k,2,1), 1), 1);
elseif y(k) == 2
mu(k,1,1) = 1 -mu(k,1,2);
mu(k,2,1) = tapas_logit(1 -tapas_sgm(mu(k,2,2), 1), 1);
end
end
else
mu(k,:,:) = mu(k-1,:,:);
pi(k,:,:) = pi(k-1,:,:);
muhat(k,:,:) = muhat(k-1,:,:);
pihat(k,:,:) = pihat(k-1,:,:);
v(k,:) = v(k-1,:);
w(k,:) = w(k-1,:);
da(k,:) = da(k-1,:);
end
end
% Remove representation priors
mu(1,:,:) = [];
pi(1,:,:) = [];
% Check validity of trajectories
if any(isnan(mu(:))) || any(isnan(pi(:)))
error('tapas:hgf:VarApproxInvalid', 'Variational approximation invalid. Parameters are in a region where model assumptions are violated.');
else
% Check for implausible jumps in trajectories
dmu = diff(mu(:,2:end));
dpi = diff(pi(:,2:end));
rmdmu = repmat(sqrt(mean(dmu.^2)),length(dmu),1);
rmdpi = repmat(sqrt(mean(dpi.^2)),length(dpi),1);
jumpTol = 16;
if any(abs(dmu(:)) > jumpTol*rmdmu(:)) || any(abs(dpi(:)) > jumpTol*rmdpi(:))
error('tapas:hgf:VarApproxInvalid', 'Variational approximation invalid. Parameters are in a region where model assumptions are violated.');
end
end
% Remove other dummy initial values
muhat(1,:,:) = [];
pihat(1,:,:) = [];
v(1,:) = [];
w(1,:) = [];
da(1,:) = [];
y(1) = [];
% Implied learning rate at the first level
mu2 = squeeze(mu(:,2,:));
mu2obs = mu2(sub2ind(size(mu2), (1:size(mu2,1))', y));
mu1hat = squeeze(muhat(:,1,:));
mu1hatobs = mu1hat(sub2ind(size(mu1hat), (1:size(mu1hat,1))', y));
upd1 = tapas_sgm(mu2obs,1) -mu1hatobs;
lr1 = upd1./da(:,1);
% Create result data structure
traj = struct;
traj.mu = mu;
traj.sa = 1./pi;
traj.muhat = muhat;
traj.sahat = 1./pihat;
traj.v = v;
traj.w = w;
traj.da = da;
% Updates with respect to prediction
traj.ud = mu -muhat;
% Psi (precision weights on prediction errors)
psi = NaN(n-1,l);
pi2 = squeeze(pi(:,2,:));
pi2obs = pi2(sub2ind(size(pi2), (1:size(pi2,1))', y));
psi(:,2) = 1./pi2obs;
for i=3:l
pihati = squeeze(pihat(:,i-1,:));
pihatiobs = pihati(sub2ind(size(pihati), (1:size(pihati,1))', y));
pii = squeeze(pi(:,i,:));
piiobs = pii(sub2ind(size(pii), (1:size(pii,1))', y));
psi(:,i) = pihatiobs./piiobs;
end
traj.psi = psi;
% Epsilons (precision-weighted prediction errors)
epsi = NaN(n-1,l);
epsi(:,2:l) = psi(:,2:l) .*da(:,1:l-1);
traj.epsi = epsi;
% Full learning rate (full weights on prediction errors)
wt = NaN(n-1,l);
wt(:,1) = lr1;
wt(:,2) = psi(:,2);
wt(:,3:l) = 1/2 *(v(:,2:l-1) *diag(ka(2:l-1))) .*psi(:,3:l);
traj.wt = wt;
% Create matrices for use by the observation model
infStates = NaN(n-1,l,b,4);
infStates(:,:,:,1) = traj.muhat;
infStates(:,:,:,2) = traj.sahat;
infStates(:,:,:,3) = traj.mu;
infStates(:,:,:,4) = traj.sa;
return;