Loading a Hidden Markov Model into the HMM Module

Once the model is generated by TANGRAM-II, the user may, finally, load it into the HMM Module. To open the HMM Module, just click in the HMM Module button, shown in Figure . This will open the first window of the HMM Module interface, which is shown in Figure .

Figure: HMM Module (a) selection button; (b) model selection interface.

[]

[]

In the first window of the HMM Module interface, the user may select which type of hidden Markov model he wants to work with. As stated previously, the HMM Module supports four types of HMMs, described below:

1. Hidden Markov Model: a regular HMM which has, associated to each state, a symbol emission probability distribution.

2. Hierarchical Gilbert Hidden Markov Model: a hierarchical HMM which has, associated to each state, a Gilbert Markov model, which is responsible for the symbol emissions. This model has a fixed batch size (number of symbols emitted between two transitions in the hidden chain) whose value is determined by the user. An example of this model is illustrated on Figure .

3. Hierarchical General Hidden Markov Model - Fixed Batch: a hierarchical HMM which has, associated to each state, a custom Markov chain, which is responsible for the symbol emissions. This model has a fixed batch size (number of symbols emitted between two transitions in the hidden chain) whose value is determined by the user. An example of this model is illustrated on Figure .

4. Hierarchical General Hidden Markov Model - Variable Batch: a hierarchical HMM which has, associated to each state, a custom Markov chain with an absorbing state, which is responsible for the symbol emissions. This model has a variable batch size (number of symbols emitted between two transitions in the hidden chain) whose value is determined by the number of transitions (in the lower-level chain) it takes to reach the absorbing state. An example of this model is illustrated in Figure .

After choosing one of these models, the HMM Module will load the chain structure designed by the user in the Model Specification Module. To see how this can be done, let's continue with the example we have been working on.

Recall that, in section , we designed a hierarchical Gilbert hidden Markov model, whose structure is illustrated in Figure . Thus, when opening the HMM Module interface, we will choose the Hierarchical Gilbert Hidden Markov Model option, as shown in Figure , and click OK. This will open the second HMM Module interface, shown in Figure . In it, the user will have to specify which of his Tangram-II model's state variables correspond to the chain's upper-level states, and which correspond to the chain's lower-level states. This is necessary in order to allow the HMM Module to correctly interpret the Markov chain created by TANGRAM-II, and, consequently, accurately extract the HMM model's structure designed in the Model Specification Module. In our example, recall that we used the state variable Markov_Chain.N to represent the upper-level chain, and the state variable On_Off_Source.Status to represent the lower-level chains. Hence, we will choose them accordingly, as illustrated in Figure , and click OK.

Figure: HMM Module (a) state variable selection interface; (b) additional parameter specification interface.

[]

[]

Our HMM's structure is now created. The only thing left for us to do is specify any additional parameters that might be required by the model we chose to work with. This will be done in the third HMM Module interface, which is illustrated in Figure . In the case of the hierarchical Gilbert HMM, these parameters are: initial probability of hidden states and the batch size. Once they are specified, all you have to do is click on the Create button, and the model you designed in Tangram-II's Model Specification Module will, finally, be loaded into the HMM Module.