Additive and nonadditive fuzzy hidden Markov models
IEEE Transactions on Fuzzy Systems
| Hi-index | 0.00 |
In this work we described three new sequence processing algorithms and a new training method for a Takagi-Sugeno fuzzy rule system. The new algorithms were applied to landmine detection and to E. coli source tracking. The first algorithm is a hidden Markov model (HMM) with sigmoidal state memberships (SHMM). The main advantages of the SHMM over the continuous HMM are stability of the discriminative training and speed. We applied the SHMM to GPR landmine detection. The second algorithm is a HMM with state memberships represented using fuzzy rule systems (FHMM). The main advantage of the FHMM over the continuous HMM is that it provides a systematic way to include expert knowledge into the model. We applied the FHMM to E. coli source tracking. In order to train the FHMM with the expectation-maximization (EM) algorithm, we developed an EM training algorithm for a Takagi-Sugeno fuzzy rule system (TSFRS). The main advantage of the new training is that it does not require target values. The third algorithm is a linguistic HMM (LHMM). The main advantage of the LHMM is that it is able to process fuzzy sequences. We applied the new model to the recognition of a basketball play called pick-and-roll. The basketball scenes were described using fuzzy spatial relations. We found that the LHMM classified the basketball scenes better than the CHMM.