mpib_memoreye/behavioural_modeling/gridsearch_jld_trial.m

%% "Fitting" to human accuary and bias data (grid search) 

% update: squircleBehave2 should now yield A at the exact same scale as in 
% RSA/RDM-mixtue (see mixvectors.m in old_dev)

function [BestParams]=gridsearch(aCueI,tCueI,sim,ob,CvetIn,granu);

% Input: 
% aCueI ccw responses (1) and cw (0) per ori (1st dim) and trial (2nd dim)
% tCueI correct responses in the rest; ccw (1) and cw (0) per ori (1st dim) and trial (2nd dim)
% sim=1 will replace your data by toy data (useful for parameter recovery
% etc.). When fitting subjects, use: gridsearch(yourAccu,yourBias,0)
% ob =1 ssqmat only based on accu
% granu is the granularity of the grid search
% Output: fitted params: [s (noise), A (squircle factor), C (cardinal precision), min SSQ] 

%% simulate toy data (replaces yourData, i.e. skip when fitting subjects)
% can run this directly with "gridsearch(1,1,1)" from command line
if sim
    s=0.2;     % noise (in memory/decision-making)
    A=1;      % Key Parameter (squircle) 0: all circle; 1: all square 
    C=0.6;      % Reduce noise near cardinal by this factor (1: off)
    makefig=0;
    makefigonlybias=1;
    [yourAccu yourBias]=squircleBehave2(s,A,C,makefig,makefigonlybias);
end

%% Estimate parameters (grid search)

% Specify parameter grid (compromise granularity <->  computation time) 
Avec=-1:granu:1;
svec=0:granu:1;
Cvec=CvetIn; % to omit Cardinal precision estimation, set Cvec=1

% allocate results mat
ssqmat=NaN(length(svec),length(Avec),length(Cvec));
nits=prod(size(ssqmat)); % this many iterations will be in total

% scan the grid
fig=0; % no plotting on iterations
fig2=0; % no plotting on iterations

for i=1:length(svec)
    for j=1:length(Avec)
        for k=1:length(Cvec)
            
            %%
            [CP_P, CP_N, AccuMix BiasMix]=squircleBehave3(svec(i),Avec(j),Cvec(k),fig,fig2);
           

SSpt=[];
for or=1:length(CP_P)
    
 %Selecting trials where the correct response for the test was CW (and
%compare with CP_P

ii=(tCueI(or,:)==1&~isnan(aCueI(or,:)));
SSpt=[SSpt (CP_P(or)-aCueI(or,ii)).^2];

 %Selecting trials where the correct response for the test was CCW (and
%compare with CP_N

iii=(tCueI(or,:)==0&~isnan(aCueI(or,:)));
SSpt=[SSpt (CP_N(or)-aCueI(or,iii)).^2];

end

ssqmat(i,j,k)=sum(SSpt);            
            
            
        end
    end
end

% find minimum
[v,loc] = min(ssqmat(:));
[ii,jj,k] = ind2sub(size(ssqmat),loc);
BestParams=[svec(ii) Avec(jj) Cvec(k) min(ssqmat(:))];

end