Viterbi Algorithm

1
Viterbi Algorithm

• Let X be a path of length L. For k ? Q and 0 ? i
  ? L, consider a path ? ending at k. If vk(i ) is
  the probability of most probable path that ends
  in state k.
• Initialize
• Recursive relation For each i 0, , L -1 and
  for each l ? Q
• The value of p (X, ?) is given by

2
Viterbi Algorithm

• By keeping pointers backwards, the most optimal
  (probable) state sequence can be found on
  backtracking.
• Start backtracking from the state where vk(L ) is
  maximum for all k ? Q.
• Predicted states by Viterbi algorithm on the
  casino example for 300 rolls of a die

Rolls 315116246446644245311321631164152133625144
54363165662656666665116645313265124563666463163666
31623264 Die FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
FFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLLLFFFFFFFFFFFFLLLLL
LLLLLLLLLLLFFFLLL Viterbi FFFFFFFFFFFFFFFFFFFFFFFF
FFFFFFFFFFFFFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLFFFFFFFF
FFFFLLLLLLLLLLLLLLLLLLLLLL Rolls
55236266666625151631222555441666566563564324364131
51346514635341112641462625335636616366646623253441
Die LLLLLLLLLLLFFFFFFFFFFFFFFFFFLLLLLLLLLLLLL
FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFLLLLLLLLLLFFF
FFFFFFFFF Viterbi LLLLLLLLLLLLFFFFFFFFFFFFFFFFFFFF
FFFFFFFFFFPFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFLLL
LLLLLLLLLLFFFFFFFF Rolls 3661661163252562462255
26525226643535333623312162536441443233516324363366
5562466662632666612355245242 Die
FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
FFFFFFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLLLLFFFFFFFFFFF
Viterbi FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
FFFFFFFFFFFFFFFFFFFFFFFFFFFFFLLLLLLLLLLLLLLLLLLLFF
FFFFFFFFF
3
The Forward Algorithm

• We want to be able to calculate the probability
  of a given sequence, (as we did for a Markov
  chain), but under HMM conditions.
• However, many different state paths can give rise
  to the same sequence X in the case of a HMM.
• So add the probabilities of all possible paths
  to get the final probability.

4
The Forward Algorithm

• Given the sequence X (x1, , xL), denote fk (i
  ) to be the probability of emitting prefix (x1,
  , xi) and eventually reaching the state ?i k
• fk (i ) is the probability of the given sequence
  up to xi requiring that ?i k
• Initial values

5
The Forward Algorithm

• Recursive relation
• Terminal value
• Unlike the Viterbi algorithm, we have sums of
  probabilities.
• So, logarithms could not be used easily to avoid
  underflow errors.
• Use exponential functions or a scaling method.

6
The Backward Algorithm

• Complementary to the forward algorithm.
• Denote bk (i ) to be the probability of emitting
  suffix (xi1, , xL) given ?i k
• Initial Values
• Recursive relation
• Terminal value

7
The Posterioir Decoding Problem

• The Viterbi algorithm finds the most probable
  path through the model given a sequence of
  symbols.
• However, in general we want to find the
  probability that the observation xi came from
  state k, given the observed sequence.
• This is called posterior probability of state k
  at step i when the emitted sequence is known.
• Posterior probability is particularly useful when
  many different paths compete for the most
  probable path with almost the same probability.
• With posterior probability, we can ask questions
  like - Does Nth base in the sequence come from a
  CpG island or not?

8
Posterioir Probability

• Posterior probability is obtained by using
  forward and backward probabilities.
• By the definition of conditional probability,
  PAB PA,B / PB
• where p(X) is the result of either forward or
  backward calculation.

9
Posterioir Probability

• The posterior probability of the die being fair
  in the casino example can be calculated for each
  roll of a die.

X-axis no. of rolls, Y-axis p (die is fair)
The shaded areas show the rolls generated by
loaded die.
10
Parameter Estimation for HMM

• All examples considered so far assume that
  transmission and emission probabilities (? in HMM
  model) are known beforehand.
• In practice, we do not know these HMM model
  parameters to begin with. (example CpG island
  case).
• If we have a set of sample sequences X1, , Xn of
  lengths L1, , Ln, (called training sequences)
  then we can construct the HMM that will best
  characterize the training sequences.
• Our goal is to find ? such that the logarithmic
  scores of the training sequences are maximized.

11
Estimation When State Sequence Known

• Assume that state sequences ?1, , ?n are known.
• In CpG island case, this corresponds to a given
  set of genomic training sequences in which CpG
  islands are already labeled.
• First scan the sequences and compute
• Akl no. of transitions from state k to l, and
• Ek(b) no. of times symbol b was emitted in
  state k.
• Then the maximum likelihood estimations are

12
Estimation When State Sequence Unknown

• Called Baum-Welch training algorithm - an
  iterative technique.
• Initialize by assigning arbitrary values to ?.
• compute the expected no. of state transitions
  from k to l using,
• then the expectations are,
• where fkj (i ) and bkj(i ) are the forward and
  backward probabilities of the sequence Xj.

13
Estimation When State Sequence Unknown

• compute the expected no. of emissions of symbol
  b in the state k using,
• Maximization Re-compute the new values for ?
  from Akl and Ek(b), as in the case of known state
  sequence.
• Repeat steps 2 and 3 until the improvement of
  is less than a given
  parameter ?.
• The Baum-Welch algorithm does not guarantee
  convergence to the global maximum, it may get
  stuck in a local maximum.
• Final solution is particularly sensitive to
  initial values.