Changes Made to the Package

Description

This page contains a listing of recent changes made to the package.

Details

1. Since we have included different classes of HMMs (see dthmm, mmglm and mmpp), it is much tidier to use an object orientated approach. This ensures that the functions across all models follow a more consistent naming convention, and also the argument list for the different model functions are more simple and consistent (see Overview). (14 Sep 2007)
2. The main tasks (model fitting, residuals, simulation, Viterbi, etc) can now be called by generic functions (see topic Overview). The package documentation has been rearranged so that these generic functions contain the documentation for all model types (e.g. see BaumWelch). (14 Sep 2007)
3. There are a number of functions, still contained in the package, that are obsolete. This is either because they do not easily fit into the current naming convention used to implement the more object orientated approach, or their argument list is slightly complicated. These functions have been grouped in the topics dthmm.obsolete and mmpp.obsolete. (14 Sep 2007)
4. There are various second level functions. For example, the model fitting is achieved by the generic BaumWelch function. However, this will call functions to do the E-step, M-step, forward and backward probabilities, and so on. At the moment, these second level functions have not been modified into an object orientated approach. It is not clear at this point whether this would be advantageous. If one went down this route, then one would probably group all of the E-step functions (for all models) under the same topic. If not, then it may be best to group all second level functions for each model under the same topic (e.g. forwardback, probhmm and Estep would be grouped together, being the second level functions for the dthmm model). (14 Sep 2007)
5. The original function called Viterbi has been renamed to Viterbihmm, and Viterbi is now a generic function. (14 Sep 2007)
6. Programming code that uses old versions of the functions should still work with this revised version of the package. However, you will get warning messages stating that certain functions are deprecated, and suggesting a possible alternative. To get a quick overview of the programming style, have a look at the examples in topic dthmm. (09 Nov 2007)

Future Development

1. The functions residuals and Viterbi need methods for objects with class mmpp.
2. A number of the original functions have names that are too general. For example forwardback calculates the forward-backward probabilities, but only for the model dthmm. The corresponding function for the mmpp model is forwardback.mmpp. It would be more consistent to attach to these original functions a dthmm suffix.
3. The for loops in the evaluation of the forward and backward equations are too slow. These should be implemented in C.
4. The demonstration examples are all for dthmm. Also need some for mmglm and mmpp.