hmmtrain.m with unknown state sequence (Baum-Welch)

I have a vector of observations Y. I do not know the state sequence of the latents. I wish to find the the transistion and emission matrices hence I want Baum-Welch.

My Min Working example is here:

T = 1000;                                       % Number of timesteps
Y= 1000+cumsum(randn(T,1));
K = 200;                                        %number of states
beta = 0.5; 
TRGUESS   = get_stateTransitionMatrix(K, beta);  %flat start model
N = 2;
EMITGUESS = (1/N) .* ones(K,N);

[TRANS,EMIS] = hmmtrain(Y,TRGUESS,EMITGUESS);

function TRGUESS = get_stateTransitionMatrix(K, beta)
TRGUESS = beta.*eye(K,K);
for i=1:K
    for j=1:K
        if (TRGUESS(i,j)==0)
            TRGUESS(i,j) = (1-beta)/(K-1);
        end
    end
end
end

On running this I get:

Error using hmmdecode (line 100)
SEQ must consist of integers between 1 and 1.

Error in hmmtrain (line 213)
            [~,logPseq,fs,bs,scale] = hmmdecode(seq,guessTR,guessE);

I fear I have misunderstood the hmmtrain documentation. Can anyone help?

thanks!

(using 2012A and all the toolboxes)

[Matlab2010]

Follow up -- just to be clear, hmmtrain.m supports the discrete case.

if you have Gaussian (or other) emissions, then you need to look elsewhere eg http://code.google.com/p/pmtk3/

Answer 1

[Sean de Wolski]

Y, your input for seqs (the first input) need to be integers ranging from 1:n that are the discrete values for training.

You can map back to the original values (whatever these integers happen to represent) later.

More

T=1000;                % Number of timesteps
Y= 1000+cumsum(randn(T,1));
[uV,~,seqs] = unique(Y); %map unique values to their indices
seqs2 = rem(seqs,24)+1;  %redefine as only 25 states
N = 5;  %states
M = 25; %24+1
A = ones(N)/N;
B = ones(N,M)/M;
[TRANS,EMIS] = hmmtrain(seqs2',A,B);

[Matlab2010]

ah. thank you! a lot clearer. However, when people say that Baum-Welch is for when you don't know the hidden states, have we not just guessed them here, for the training?

thank you!!

[Sean de Wolski]

Guessing means we didn't need to know :)

I don't have time to look into your second question today.

Answer 2

[imene s]

please i need this fuction hmmtrain

[Sean de Wolski]

It's in the Statistics Toolbox which you or your company/university will need to purchase.

Answer 3

[amani]

Any thoughts about the default prior vector (pi) used in hmmtrain.m ??