Scribd
Upload
32 views2 pages

Hidden Markov Models Common Probabilities HMM Diagram

This document summarizes key concepts and equations for hidden Markov models (HMMs). It includes: 1) An HMM diagram showing latent and observed states over time, with transition probabilities (aij) and observation probabilities (bi(ot)). 2) Equations for decomposing the probability of being in state si at time t and sj at t+1 (ξt(i,j)) in terms of backward (α), forward (β), and transition (aij) probabilities. 3) A recursive equation for calculating the forward probability (βt(i)) in terms of β probabilities at the next time step. 4) An equation for the probability of being in state sj

Uploaded by

Nidhal Selmi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
32 views2 pages

Hidden Markov Models Common Probabilities HMM Diagram

This document summarizes key concepts and equations for hidden Markov models (HMMs). It includes: 1) An HMM diagram showing latent and observed states over time, with transition probabilities (aij) and observation probabilities (bi(ot)). 2) Equations for decomposing the probability of being in state si at time t and sj at t+1 (ξt(i,j)) in terms of backward (α), forward (β), and transition (aij) probabilities. 3) A recursive equation for calculating the forward probability (βt(i)) in terms of β probabilities at the next time step. 4) An equation for the probability of being in state sj

Uploaded by

Nidhal Selmi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Nidhal Selmi, Hidden Markov Models nselmi@asu.

edu
Tanishka Singh common probabilities Spring 2019

HMM Diagram:
si sj
aij
Latent Q: { q1 ... qt qt+1 qt+2 ... qT }

bi (ot ) bj (ot+1 )

Observed O: { o1 ... ot ot+1 ot+2 ... oT }

o1:t ot+2:T

ξt(i, j):Probability to be in state si at t, and si+1 at t + 1


Let oi:j denote the sequence of j − i + 1 observations (oi , oi+1 , ..., oj−1 , oj ).
λ is ommitted since the model is a given for all the probabilities presented.

αt (i) =P (o1:t , qt = si ) Backward prob of observations 1...t and state j after


βt+1 (j) =P (ot+2:T |qt+1 = sj ) Forward prob of observations t+2...T given state j

To decompose ξ into a product of backward, forward and transition probabilities,


we use a combination of Bayes theorem (1,4,6) and d-separation (5,8).

P (qt = si , qt+1 = sj , O)
ξt (i, j) = P (qt = si , qt+1 = sj |O) = (1)
P (O)
P (O)ξt (i, j) = P (qt = si , qt+1 = sj , O) = P (qt = si , qt+1 = sj , o1:t , ot+1:T ) (2)
P (O)ξt (i, j) = P (qt = si , qt+1 = sj , o1:t , ot+1:T ) (3)
P (O)ξt (i, j) = P (o1:t , qt = si , ot+1:T , qt+1 = sj ) (4)
P (O)ξt (i, j) = P (o1:t , qt = si ) P (ot+1:T , qt+1 = sj |o1:t , qt = si ) (5)
| {z }
αt (i)

P (O)ξt (i, j) = αt (i)P (ot+1:T , qt+1 = sj |qt = si ) (6)


P (O)ξt (i, j) = αt (i)P (ot+1:T |qt+1 = sj , qt = si )P (qt+1 = sj |qt = si ) (7)
| {z }
aij

P (O)ξt (i, j) = αt (i)aij P (ot+1 , ot+2:T |qt+1 = sj , qt = si ) (8)


P (O)ξt (i, j) = αt (i)aij P (ot+1 |qt+1 = sj ) P (ot+2:T |qt+1 = sj ) (9)
| {z }| {z }
bj (ot+1 ) βt+1 (j)

αt (i)aij bj (ot+1 )βt+1 (j)


ξt (i, j) = (10)
P (O|λ)

1
Nidhal Selmi, Hidden Markov Models [email protected]
Tanishka Singh common probabilities Spring 2019

βt(i) Forward probability recursively in terms of βt+1(j)

βt (i) = P (ot+1:T |qt = si )


βt (i) = P (ot+1 , ot+2:T |qt = si ) (separate ot+1 given qt )
βt (i) = P (ot+1 |qt = si )P (ot+2:T |qt = si )

Marginalizing over all next possible states j (with transitions and observations) given state i.
N
X
βt (i) = P (qt+1 = sj |qt = si )P (ot+1 |qt+1 = sj )P (ot+2:T |qt+1 = sj )
j=1
N
X
βt (i) = aij bj (ot+1 )βt+1 (j)
j=1

γt(j):Probability to be in state sj given observations O

P (qt = sj , O)
γt (j) = P (qt = sj |O) = (Bayes)
P (O)
P (O)γt (j) = P (qt = sj , O)
P (O)γt (j) = P (qt = sj , o1:t , ot+1:T ) (split O)
P (O)γt (j) = P (o1:t , qt = sj , ot+1:T )
P (O)γt (j) = P (o1:t , qt = sj )P (ot+1:T |o1:t , qt = sj ) (Bayes)
P (O)γt (j) = P (o1:t , qt = sj )P (ot+1:T |qt = sj ) (separate o1:t )
αt (j)βt (j)
γt (j) =
P (O|λ)

References
[1] L. R. Rabiner, ”A tutorial on hidden Markov models and selected applications in speech
recognition,” in Proceedings of the IEEE, vol. 77, no. 2, pp. 257-286, Feb. 1989.

You might also like