08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

a ij<br />

transition probability from state i to state j<br />

b j (O t+1 ) probability <strong>of</strong> O t+1 given that the HMM is in state j at time t + 1<br />

α t (i)<br />

β t+1 (j)<br />

δ(i, j)<br />

s t (i)<br />

p(O)<br />

probability <strong>of</strong> seeing O 0 O 1 · · · O t and ending in state i at time t<br />

probability <strong>of</strong> seeing the tail <strong>of</strong> the sequence O t+2 O t+3 · · · O T given state j<br />

at time t + 1<br />

probability <strong>of</strong> going from state i to state j at time t given the sequence<br />

<strong>of</strong> outputs O<br />

probability <strong>of</strong> being in state i at time t given the sequence <strong>of</strong> outputs O<br />

probability <strong>of</strong> output sequence O<br />

The probability <strong>of</strong> being in state i at time t is given by<br />

n∑<br />

s t (i) = δ t (i, j).<br />

j=1<br />

Note that δ t (i, j) is the probability <strong>of</strong> being in state i at time t given O 0 O 1 O 2 · · · O t but<br />

it is not the probability <strong>of</strong> being in state i at time t given O since it does not take into<br />

account the remainder <strong>of</strong> the sequence O. Summing s t (i) over all time periods gives the<br />

expected number <strong>of</strong> times state i is visited and the sum <strong>of</strong> δ t (i, j) over all time periods<br />

gives the expected number <strong>of</strong> times edge i to j is traversed.<br />

Given estimates <strong>of</strong> the HMM parameters a i,j and b j (O k ), we can calculate by the above<br />

formulas estimates for<br />

1. ∑ T −1<br />

i=1 s t(i), the expected number <strong>of</strong> times state i is visited and departed from<br />

2. ∑ T −1<br />

i=1 δ t(i, j), the expected number <strong>of</strong> transitions from state i to state j<br />

Using these estimates we can obtain new estimates <strong>of</strong> the HMM parameters<br />

a ij =<br />

expected number <strong>of</strong> transitions from state i to state j<br />

expected number <strong>of</strong> transitions out <strong>of</strong> state i<br />

b j (O k ) = expected number <strong>of</strong> times in state j observing symbol O k<br />

expected number <strong>of</strong> times in state j<br />

307<br />

=<br />

∑ T −1<br />

t=1 δ t(i, j)<br />

∑ T −1<br />

t=1 s t(i)<br />

=<br />

T∑<br />

−1<br />

t=1<br />

subject to<br />

O t=O k<br />

s t (j)<br />

∑ T −1<br />

t=1 s t(j)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!