The main algorithms used in the seqhmm package

Transcription

1 The main algorithms used in the seqhmm package Jouni Helske University of Jyväskylä, Finland May 9, Introduction This vignette contains the descriptions of the main algorithms used in the seqhmm (Helske and Helske, 2017) package. First, a forward-backward algorithm is presented, followed by a Viterbi algorithm, and the derivations of the gradients for the numerical optimisation routines. 2 Forward Backward Algorithm Following Rabiner (1989), the forward variable α it (s) P (y i1,..., y it, z t s M) is the joint probability of partial observation sequences for subject i until time t and the hidden state s at time t given the model M. Let us denote b s (y it ) b s (y it1 ) b s (y itc ), the joint emission probability of observations at time t in channels 1,..., C given hidden state s. The forward variable can be solved recursively for subject i 1,..., N: α i1 (s) π s b s (y i1 ) 2. Recursion: For [ t 1,..., T 1, and r 1,..., S, compute S α i(t+1) (r) α it(s)a sr b r (y i(t+1) ) 3. Termination: Compute the likelihood P (Y i M) S α it (s) The backward variable β it (s) P (y i(t+1),..., y it z t s, M) is the joint probability of the partial observation sequence after time t and hidden state s at time t given the model M. For subject i 1,..., N, the backward variable can be computed as 1

2 1. Initialization: For s 1,..., S, set β it (s) 1 2. Recursion: [ For t T 1,..., 1, and s 1,..., S, compute S β it (s) r1 a sr b s (y i(t+1) )β i(t+1) (r) In practice the forward-backward algorithm is prone to numerical instabilities. Typically we scale the forward and backward probabilities, as follows (Rabiner, 1989). For subject i 1,..., N, α i1 (s) π s b s (y i1 ), c t 1/ α i1 (s), ˆα i1 (s) c t α i1 (s) (1) 2. Recursion: For t 1,..., T 1, and r 1,..., S, compute [ S α i(t+1) (r) α it (s)a sr b r (y i(t+1) ) c t 1/ α i(t+1) (s), ˆα i(t+1) (s) c t α i(t+1) (s) (2) 3. Termination: Compute the log-likelihood logp (Y i M) T t1 c t The scaling factors c t from the forward algorithm are commonly used to scale also the backward variables, although other scaling schemes are possible as well. In seqhmm, the scaled backward variables for subject i 1,..., N are computed as ˆβ it (s) 1 (3) 2. Recursion: For t T 1,..., 1, and r 1,..., S, compute [ S β it (s) a sr b s (y i(t+1) )β i(t+1) (r), r1 ˆβ it (s) c t β it (s) (4) 2

3 Most of the times this scaling method described works well, but in some illconditioned cases it is possible that the default scaling still produces underflow in backward algorithm. For these cases, seqhmm also supports the computation of the forward and backward variables in log-space. Although numerically more stable, the algorithm is somewhat slower due repeated use of log-sum-exp trick. 3 Viterbi Algorithm We define the score δ it (s) max P (z i1 z it s, y i1 y it M), z i1z i2 z i(t 1) which is the highest probability of the hidden state sequence up to time t ending in state s. By induction we have [ δ i(t+1) (r) max δ it (s)a sr b r (y i(t+1) ). (5) s We collect the arguments maximizing Equation 5 in an array ψ it (r) to keep track of the best hidden state sequence. The full Viterbi algorithm can be stated as follows: 1. Initialization δ i1 (s) π s b s (y i1 ), s 1,..., S ψ i1 (s) 0 2. Recursion δ it (r) max,...,s (δ i(t 1) (s)a sr )b h (y it ), ψ it (s) arg max,...,s (δ i(t 1) (s)a sr ), s 1,..., S; t 2,..., T 3. Termination ˆP max,...,s (δ it (s)) ẑ it arg max,...,s (δ it (s)) 4. Sequence backtracking ẑ it ψ i(t+1) (ŝ i(t+1) ), t T 1,..., 1. To avoid numerical underflow due to multiplying many small probabilities, the Viterbi algorithm can be straighforwardly computed in log space, i.e., calculating log(δ it (s)). 4 Gradients Following Levinson, Rabiner, and Sondhi (1983), by using the scaled forward and backward variables we have π s b s (y i1 ) ˆβ i1 (s), 3

4 and T 1 ˆα it (s)b r (y i(t+1) ) a ˆβ i(t+1) (r), sr t1 b rc (m) t:y itcm α 1t (s)a sr ˆβi(t+1) (r) + I(y i1c m)π r ˆβi1 (r). In the direct numerical optimization algorithms used by seqhmm, the model is parameterised using unconstrained parameters π s, a sr, b rc(m) such that a sr exp(a sr)/ S k1 exp(a sk), and similarly for emission and initial probabilities. This leads to and π s a sr b rc(m) log P (Y i M) π s π s (1 π s ) log P (Y i M) a sr a sr (1 a sr ), log P (Y i M) b rc (m)(1 b rc (m)). b rc (m) 4.1 MHMM case For mixture HMM with K clusters, we define a full model with S S S K states in a block form with π i (w i1 π 1,..., w ik π K ), where π k, k 1,..., K is the vector of initial probabilities for the submodel M k and w ik exp(x i γ k)/(1 + K j2 exp(x i γ j), with γ 1 0. First note that the log-likelihood of the HMM for ith subject can be written as P (Y i M) α t (s)a sr b r (y i(t+1) )β t+1 (r), r1 for any t 1,..., T 1. Thus for t 1 we have P (Y i M) α 1 (s)a sr b r (y i2 )β 2 (r) r1 α 1 (s) a sr b r (y i2 )β 2 (r) r1 α 1 (s)β 1 (s) π is b s (y i1 )β 1 (s). (6) 4

5 Therefore the gradients for the unconstrained parameters πs k cluster are given as π k s For γ k, the gradients are of form log P (Y i M k ) πs k πs k (1 πs k )w ik. Now if state s belongs to cluster k, we have of the kth b s (y i1 ) ˆβ 1 (s) π is. (7) π is π k s exp(x i γ k) γ K k j1 exp(x i γ j)) π k s x i w ik (1 w ik ), (8) and π is π h s x i w ih w ik, otherwise, where h is the index of cluster containing the state s. References Satu Helske and Jouni Helske. Mixture hidden Markov models for sequence data: the seqhmm package in R. Preprint ArXiv: , URL S. E. Levinson, L. R. Rabiner, and M. M. Sondhi. An introduction to the application of the theory of probabilistic functions of a Markov process to automatic speech recognition. Bell System Technical Journal, 62(4): , ISSN doi: /j tb03114.x. URL Lawrence Rabiner. A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2): , doi: /