4www.intechopen.com observations, these techniques can be used to choose the HMM that most likely generated the sequence of observations and, in the process, also characterize the associated probability of error (for the given sequence of observations). However, in order to measure the classification capability of the classifier before making any observations, one needs to compute the ap r i o r i probability that the classifier makes an incorrect decision for any of the possible sequences of observations. Enumerating all possible sequences of a given length (in order to evaluate their contribution to the probability of error) is prohibitively expensive for long sequences; thus, we describe ways to avoid this computational complexity and obtain an upper bound on the probability that the classifier makes an error without having to enumerate all possible output sequences. Specifically, we present a constructive approach that bounds the probability of error as a function of the observation step. We also discuss necessary and sufficient conditions for this bound on the probability of error to go to zero as the number of observations increases. After obtaining bounds on the probability of erroneous classification, we consider the additional challenge that the observed sequence is corrupted, due to noise coming from sensor malfunctions, communication limitations, or other adversarial conditions. For example, depending on the underlying application, the information that the sensors provide may be corrupted due to inaccurate measurements, limited resolution, or degraded sensor performance (due to aging or hardware failures). We consider unreliable sensors that may cause outputs to be deleted, inserted, substituted or transposed with certain known probabilities. Under such sensor malfunctions, the length of the observed sequence will generally not equal the length of the output sequence and, in fact, several output sequences may correspond to a given observed sequence. Thus, one would need to first identify all possible state sequences and the probabilities with which they agree with both the underlying model and the observations (after allowing, of course, for sensor failures). In particular, if symbols in the output sequence can be deleted, there may be an infinite number of output sequences that agree with a given observed sequence, which makes the standard forward algorithm inapplicable for classification. This inability of the standard forward algorithm can be overcome via an iterative algorithm that allows us to efficiently compute the probability that a certain model matches the observed sequence: each time a new observation is made, the algorithm simply updates the information it keeps track of and outputs on demand the probability that a given model has produced the sequence observed so far. The iterative algorithm we describe relates to (and generalizes) iterative algorithms for the evaluation problem in HMMs (Rabiner, 1989), the parsing problem in probabilistic automata (PA) (Fu, 1982), (Vidal et al., 2005), a...