So-called hidden semi-Markov models (HSMMs) are a flexible generalization of HMMs to a semi-Markovian state process which is motivated by the fact that for homogeneous HMMs, the time spent in a hidden state, also called the state dwell time or sojourn time is necessarily geometrically distributed as a consequence of the Markov assumption. HSMMs are designed to mitigate this often unrealistic assumption by allowing for arbitrary distributions on the positive integers to be estimated for the state dwell time. Inference in such models becomes more involved, but Langrock and Zucchini (2011) showed that HSMMs can be estimated conveniently via approximating them by HMMs with an extended state space. Each state of the HSMMs is represented by a state aggregate of several states and the transition probabilities within each aggregate are designed carefully to represent the chosen dwell-time distribution. For more details see Langrock and Zucchini (2011) or Zucchini, MacDonald, and Langrock (2016). Due to this approximate inference procedure, such models can again be fitted by numerically maximizing the (approximate) likelihood which can be evaluated using the forward algorithm.
We begin by considering homogeneous HSMMs. In such models, each state has an associated state dwell-time distribution. The transition probability matrix of a regular HMM is replaced by these distributions and the conditional transition probabilities given the state is left.
Here we choose the simplest case of dwell times that are shifted Poisson distributed. We have to specify the Poisson mean for each state, the conditional transition probability matrix called \(\Omega\) and the parameters of the state-dependent process.
lambda = c(7, 4, 4)
omega = matrix(c(0, 0.7, 0.3,
0.5, 0, 0.5,
0.7, 0.3, 0), nrow = 3, byrow = TRUE)
mu = c(10, 40, 100)
sigma = c(5, 20, 50)
color = c("orange", "deepskyblue", "seagreen2")
curve(dnorm(x, mu[1], sigma[1]), lwd = 2, col = color[1], bty = "n",
xlab = "x", ylab = "density", xlim = c(0, 150), n = 300)
curve(dnorm(x, mu[2], sigma[2]), lwd = 2, col = color[2], add = T)
curve(dnorm(x, mu[3], sigma[3]), lwd = 2, col = color[3], add = T)
We simulate data by drawing dwell times from the dwell-time distribution of the current state and then draw the next state using the conditional transition probabilities. The state-dependent process is drawn conditional on the current state.
set.seed(123)
k = 50 # number of stays
s = rep(NA, k)
s[1] = sample(1:3, 1) # uniform initial distribution
staylength = rpois(1, lambda[s[1]]) + 1 # drawing dwell time from shifted Poisson
C = rep(s[1], staylength)
x = rnorm(staylength, mu[s[1]], sigma[s[1]])
for(t in 2:k){
# conditionally drawing state
s[t] = sample(c(1:3)[-s[t-1]], 1, prob = omega[s[t-1], -s[t-1]])
staylength = rpois(1, lambda[s[t]]) + 1 # drawing dwell time from shifted Poisson
C = c(C, rep(s[t], staylength))
x = c(x, rnorm(staylength, mu[s[t]], sigma[s[t]]))
}
plot(x, pch = 20, col = color[C], bty = "n")
legend("topright", col = color, pch = 20,
legend = paste("state", 1:3), box.lwd = 0)
We now write the negative log-likelihood function for an
approximating HMM. As a semi-Markov chain is specified in terms of
state-specific dwell-time distributions and conditional transition
probabilities given that the current state is left, we have to compute
both (here called dm
and omege
). The
transition probability matrix of the approxmiating HMM can then be
computed by the function tpm_hsmm()
where the exact
procedure is detailed by Langrock and Zucchini
(2011). We need the extra
argument agsizes
to specify the aggregate sizes that should
be used to approximate the dwell-time distributions. These should be
chosen such that most of the support of the state-specific dwell-time
distributions is covered.
mllk = function(theta.star, x, N, agsizes){
mu = theta.star[1:N]
sigma = exp(theta.star[N+1:N])
lambda = exp(theta.star[2*N+1:N])
if(N>2){
# this is a bit complicated as we need the diagonal elements to be zero
omega = matrix(0,N,N)
omega[!diag(N)] = as.vector(t(matrix(c(rep(1,N),
exp(theta.star[3*N+1:(N*(N-2))])),N,N-1)))
omega = t(omega)/apply(omega,2,sum)
} else{ omega = matrix(c(0,1,1,0),2,2) }
dm = list() # list of dwell-time distributions
for(j in 1:N){ dm[[j]] = dpois(1:agsizes[j]-1, lambda[j]) } # shifted Poisson
Gamma = LaMa::tpm_hsmm(omega, dm)
delta = LaMa::stationary(Gamma)
allprobs = matrix(1, length(x), N)
ind = which(!is.na(x))
for(j in 1:N){
allprobs[ind,j] = dnorm(x[ind], mu[j], sigma[j])
}
-LaMa::forward_s(delta, Gamma, allprobs, agsizes)
}
# intial values
theta.star = c(10, 40, 100, log(c(5, 20, 50)), # state-dependent
log(c(7,4,4)), # dwell time means
rep(0, 3)) # omega
agsizes = qpois(0.95, lambda)+1
t1 = Sys.time()
mod = nlm(mllk, theta.star, x = x, N = 3, agsizes = agsizes, stepmax = 2)
Sys.time()-t1
#> Time difference of 0.2369659 secs
HSMMs are rather slow (even using C++) as we translate the additional model complexity into a higher computational overhead (31 states here).
N = 3
(mu = mod$estimate[1:N])
#> [1] 10.16569 39.06161 107.66034
(sigma = exp(mod$estimate[N+1:N]))
#> [1] 4.78882 19.35639 48.56115
(lambda = exp(mod$estimate[2*N+1:N]))
#> [1] 6.942983 4.595469 3.354765
omega = matrix(0,N,N)
omega[!diag(N)] = as.vector(t(matrix(c(rep(1,N),
exp(mod$estimate[3*N+1:(N*(N-2))])),N,N-1)))
omega = t(omega)/apply(omega,2,sum)
omega
#> [,1] [,2] [,3]
#> [1,] 0.0000000 0.5541031 0.4458969
#> [2,] 0.5040938 0.0000000 0.4959062
#> [3,] 0.6654703 0.3345297 0.0000000
We now want to briefly show the analysis of a real data set using
hidden semi-Markov models. For this purpose we use the movement track of
an Arctic muskox contained in the R
package
PHSMM
. Originally these data where collected by Beumer et al. (2020) and have already been
analyzed by Pohle, Adam, and Beumer (2022).
# install.packages("PHSMM")
data = PHSMM::muskox[1:1000,] # only using first 1000 observations for speed
head(data)
#> date tday x y step
#> 88273 2013-10-12 15 513299.2 8264867 17.998874
#> 88274 2013-10-12 16 513283.4 8264875 8.214733
#> 88275 2013-10-12 17 513284.3 8264883 7.205098
#> 88276 2013-10-12 18 513280.4 8264877 53.378332
#> 88277 2013-10-12 19 513252.0 8264922 719.242687
#> 88278 2013-10-12 20 513386.7 8265629 10.797127
As these data have already been preprossed, we can immediately write the negative log-likelihood function. When modeling the dwell-time distribution of real processes, it is typically advisable to use a more flexible distribution than the shifted Poisson distribution, as the latter cannot account for overdispersion. Here, we will employ the shifted negative binomial distribution that yields the Poisson distribution as a special case for the dispersion parameter equal to zero. The state-dependent step lengths are modeled by gamma distributions, where we reparametrize the gamma distribution in terms of its mean and standard deviation as opposed to shape and scale for better interpretability.
mllk_muskox = function(theta.star, step, N, agsizes){
# parameter transformation from working to natural
mu = exp(theta.star[1:N]) # step mean
sigma = exp(theta.star[N+1:N]) # step standard deviation
mu_dwell = exp(theta.star[2*N+1:N]) # dwell time mean
phi = exp(theta.star[3*N+1:N]) # dwell time dispersion
if(N>2){
# conditional transition probability matrix
omega = matrix(0,N,N)
omega[!diag(N)] = as.vector(t(matrix(c(rep(1,N),
exp(theta.star[4*N+1:(N*(N-2))])),N,N-1)))
omega = t(omega)/apply(omega,2,sum)
} else{ omega = matrix(c(0,1,1,0),2,2) }
dm = list() # list of dwell-time distributions
for(j in 1:N){
# R allows to parametrize by mean and size where size = 1/dispersion
dm[[j]] = dnbinom(1:agsizes[j]-1, mu=mu_dwell[j], size=1/phi[j])
}
Gamma = LaMa::tpm_hsmm(omega, dm)
delta = LaMa::stationary(Gamma)
allprobs = matrix(1, length(step), N)
ind = which(!is.na(step))
for(j in 1:N){
# we reparametrise the gamma distribution in terms of mean and sd
allprobs[ind,j] = dgamma(step[ind], shape = mu[j]^2 / sigma[j]^2,
scale = sigma[j]^2 / mu[j])
}
-LaMa::forward_s(delta, Gamma, allprobs, agsizes)
}
# intial values
theta.star = c(log(c(4, 50, 300, 4, 50, 300)), # state-dependent mean and sd
log(c(3,3,5)), # dwell time means
log(c(0.01, 0.01, 0.01)), # dwell time dispersion
rep(0, 3)) # omega
agsizes = c(11,11,14)
t1 = Sys.time()
mod_muskox = nlm(mllk_muskox, theta.star, step=data$step, N=3,
agsizes=agsizes,iterlim = 500)
Sys.time()-t1
#> Time difference of 1.919788 secs
We retransform the parameters for interpretation
theta.star = mod_muskox$estimate; N = 3
(mu = exp(theta.star[1:N])) # step mean
#> [1] 4.408109 55.515891 306.504787
(sigma = exp(theta.star[N+1:N])) # step standard deviation
#> [1] 3.148127 50.337602 331.539171
(mu_dwell = exp(theta.star[2*N+1:N])) # dwell time mean
#> [1] 2.544975 2.660335 5.541752
(phi = exp(theta.star[3*N+1:N])) # dwell time dispersion
#> [1] 3.643785e-05 3.751638e-02 1.836526e-09
omega = matrix(0,N,N)
omega[!diag(N)] = as.vector(t(matrix(c(rep(1,N),
exp(theta.star[4*N+1:(N*(N-2))])),N,N-1)))
omega = t(omega)/apply(omega,2,sum)
omega
#> [,1] [,2] [,3]
#> [1,] 0.0000000 0.6865910 3.134090e-01
#> [2,] 1.0000000 0.0000000 2.959219e-08
#> [3,] 0.8210834 0.1789166 0.000000e+00
In this case the Poisson distribution would have been suffficiently flexible, as all dispersion parameters were estimated very close to zero. We can easily visualize the estimated state-specific dwell-time distributions: