genmarkov.m

the matlab code of ginisvm - an svm implementation good at probability regression.

function [trainx,testx, trainlabel, testlabel] = genmarkov(trmatrix,weights,param,dim,maxtrn,maxtst)
%--------------------------------------------------------------------------------------------------
%function [trainx,testx] = genmarkov(ntrain,ntest,trmatrix,weights,param,dim)
% Generates training and test data based on some 
% hidden distribution and markovian property
% using M previous states. This is useful in generating
% time series data and for speech waveforms in
% a controlled fashion. 
%
% ntrain -> number of training data
% ntest -> number of test data
% trmatrix -> transition matrix
% param -> parameter matrix for the distribution
% ( {mean, variance}  for each mixture)
% dim -> dimension of the data to be generated
% weights -> for each state the mixture probability
% 
%--------------------------------------------------------------------------------------------------

global poption;
global dynamics;
global startstate;
global finalstate;
global header;

% Start with a zero state for the random 
% number generator
[S,S] = size(trmatrix);
state = rand(2,dim,S)*1e+09;
[S,M] = size(weights);

% Initial distribution = uniform distribution
start = ceil(rand(1,1)*(S));

% Get the distribution based on the transition
% probability
% 0|------1|---2|------3|---------------4|--5|
% 
% generate random numbers before so that the 
% distribution settles down to a steady state 
% value. Therefore scramble it with some initial
% distribution.

prevstate=start;

% Generate a temporary transition matrix to
% scramble the data
temptr = 1/S*ones(S,S);

for i = 1:header,
   numberline = tril(ones(S,S))*temptr(:,prevstate);
   temp = (find(numberline>=rand(1,1)));
   prevstate = temp(1);
   clear temp;
   clear numberline;
   % Once u have chosen the next state now 
   % generate a feature vector from it
   % assuming no correlation between parameters.
   % We dont have different random number generators
   % therefore we have mean and variance.
   % Or we have to store the state of the random number
   % generator for each state.

   % Now choose a mixture
   numberline = weights(prevstate,:)*tril(ones(M,M));
   temp = (find(numberline>rand(1,1)));
   mixture = temp(1);
   clear temp;
   clear numberline;

   for d = 1:dim,
      % number of dimension
      randn('state',state(:,d,prevstate));
      randn(1,1);
      state(:,d,prevstate) = randn('state');
   end;
end;

% If the plot option is on
if poption == 1,
   figure;
   hold on;
end;

% Generate training data. Now one can start 
% from the start state and generate data according
% to the given transition matrix.
prevstate = startstate;

stopgen = 0;
ntrain = 1;
while stopgen == 0 & ntrain <= maxtrn,
   numberline = tril(ones(S,S))*trmatrix(:,prevstate);
   temp = (find(numberline>=rand(1,1)));

   prevstate = temp(1);
   if prevstate == finalstate,
      stopgen = 1;
   end;
   clear temp;
   clear numberline;
   % Once u have chosen the next state now 
   % generate a feature vector from it
   % assuming no correlation between parameters.
   % We dont have different random number generators
   % therefore we have mean and variance.
   % Or we have to store the state of the random number
   % generator for each state.

   % Now choose a mixture
   numberline = weights(prevstate,:)*tril(ones(M,M));
   temp = (find(numberline>rand(1,1)));
   mixture = temp(1);
   clear temp;
   clear numberline;

   for d = 1:dim,
      % number of dimension
      randn('state',state(:,d,prevstate));
      trainx(ntrain,d) = param(prevstate,2*(mixture-1)*(dim)+ 2*(d-1) +1) + ...
                    + param(prevstate,2*(mixture-1)*(dim)+ 2*(d-1) +2)*randn(1,1);
     state(:,d,prevstate) = randn('state');
   end;
   trainlabel(ntrain) = prevstate;
   ntrain = ntrain + 1;
   if poption == 1,
      if trainlabel(ntrain) == 1,
         plot(trainx(ntrain,1),trainx(ntrain,2),'r+');
      else
         plot(trainx(ntrain,1),trainx(ntrain,2),'b*');
      end;
   end;
end;

% Now process the test data
if poption == 1,
   figure;
   hold on;
end;
% Generate test data using the initial state.
prevstate = startstate;
stopgen = 0;
ntest = 1;
while stopgen == 0 & ntest <= maxtst,
   numberline = tril(ones(S,S))*trmatrix(:,prevstate);
   temp = (find(numberline>=rand(1,1)));
   prevstate = temp(1);
   if prevstate == finalstate,
      stopgen = 1;
   end;
   clear temp;
   clear numberline;
   % Once u have chosen the next state now 
   % generate a feature vector from it
   % assuming no correlation between parameters.
   % We dont have different random number generators
   % therefore we have mean and variance.
   % Or we have to store the state of the random number
   % generator for each state.

   % Now choose a mixture
   numberline = weights(prevstate,:)*tril(ones(M,M));
   temp = (find(numberline>rand(1,1)));
   mixture = temp(1);
   clear temp;
   clear numberline;

   for d = 1:dim,
      % number of dimension
      randn('state',state(:,d,prevstate));
      testx(ntest,d) = param(prevstate,2*(mixture-1)*(dim)+ 2*(d-1) +1) + ...
                    + param(prevstate,2*(mixture-1)*(dim)+ 2*(d-1) +2)*randn(1,1);
     state(:,d,prevstate) = randn('state');
   end;
   testlabel(ntest) = prevstate;
   ntest = ntest + 1;
   if poption == 1,
      if testlabel(ntest) == 1,
         plot(testx(ntest,1),testx(ntest,2),'r+');
      else
         plot(testx(ntest,1),testx(ntest,2),'b*');
      end;
   end;
end;