4. Sequence learning
• Many interesting application or advanced pattern are correlated
with previous data.
• sliding window based algorithm
• It could apply original machine learning algorithm easily
• How to control window length?
• Too small gives poor performance
• Too big is computationally unfeasible
11. Problem definition
• Evaluation problem
• Given a HMM, finding the probability of observation sequence.
• Decoding problem
• Given a HMM, finding the sequence of hidden states that most probably
generated an observation sequence.
• Learning problem
• Given a observation sequence and a HMM with initial parameter, let the
HMM could evaluate the observation sequence as maximal probability as
possible
• If we provide the related sequence of hidden states, HMM can be trained
such as supervised learning.
17. Unsupervised learning
Baum-Welch algorithm
Baum-Welch is one kind of maximum likelihood algorithm,
it doesn’t guarantee global maximum
ξt(i,j)= HMM0
t Si t+1 Sj
HMM0
observation sequence
γt(i)= HMM0
t Si
T(Si,Sj) = E(Ok,Sj) =
HMM0 HMM1
Evaluation problem observation sequence,
HMM1 HMM0 probability
18. Calculate
alpha( by forward algorithm)
beta( by backward algorithm)
Calculate Eplison
( transaction probability
from state i to state j in time t )
Calculate Gamma
(sum of transaction probability
at state i in time t )
Coverage
No
Begin
Yes
Finish
Expectation Step
Maximization Step