function  [F, error_rate, diff_xhist] = split_classify(input_vec, method, split_point, analysis_window, do_plot);

% split_classify( INPUT_VECTOR, METHOD, SPLIT_POINT, ANALYSIS_WINDOW, DO_PLOT);
%
% Trains on the first 1:SPLIT_POINT actions using history vectors of size 
% ANALYSIS_WINDOW+2 assuming that METHOD is the trial type that is being run. 
% Takes the following arguments:   
% 
% Required:
%   INPUT_VECTOR - The vector of choices given by the user (e.g. c1(:,1)')
%
%   METHOD - Must be either 1, 2, or 3.  Will specify the form of the reward
%   schedule presented under this trial
% 
%   SPLIT_POINT - Train on trials up to SPLIT_POINT and test/predict on trials 
%   after SPLIT_POINT
%
%   ANALYSIS_WINDOW - The size of the reward history to use.  For example,
%   if we want to include the past three rewards in the history vector,
%   then ANALYSIS_WINDOW will be 3.
%
%   DO_PLOT - Set this to 1 if you want to see the results of the model
%   trajectories plotted with the user history.  Set to 0 if no plotting
%   output is desired.
%
% The output arguments are [F, ERROR_RATE, DIFF_XHIST].  F is the structure
% containing the LDA parameters.  ERROR_RATE is the classification error
% rate, and DIFF_XHIST is the difference between model final allocation and
% subject final allocation.
%
% Examples:
%   For the examples described in our project
% 
% >> [F, error_rate, diff_xhist] = split_classify(c1(:,1)', 1, 120, 5, 0);

if size(input_vec,2)<size(input_vec,1), input_vec = input_vec'; end;

[X, labels] = generate_histories(input_vec, method, analysis_window);

F = lda(X(1:split_point,:), labels(1:split_point)+1);

output_vec = input_vec;
switch_or_not = zeros(1, length(input_vec));

for i=split_point+1:length(input_vec),
    reward_hist = play_game(output_vec(1:i-1), 1);
    Xtest = [fliplr(reward_hist(i-analysis_window:end)) sum(reward_hist) i];
    [C, P] = classify(F, Xtest);
    switch_or_not(i) = floor(rand(1)+P(2));
    output_vec(i) = xor(output_vec(i-1), switch_or_not(i));
end;

C = classify(F, X(split_point+1:end, :))-1;
errors = length(find(C~=labels(split_point+1:end)));
error_rate = errors/length(C);

new_choice_vec = output_vec;
xhist1 = get_xhist(input_vec);
xhist2 = get_xhist(new_choice_vec);
diff_xhist = abs(xhist1(end)-xhist2(end));
if do_plot,
figure(1);
h1= plot(xhist2, 'r'); hold on; h2 = plot(xhist1);
ylabel('Allocation to A');
set(h1, 'linewidth', 2);
set(h2, 'linewidth', 2);
legend('Predicted', 'Actual');
ylim([0 1]);
end;