Here you will find daily news and tutorials about R, contributed by over 573 bloggers.
There are many ways to follow us - By e-mail:On Facebook: If you are an R blogger yourself you are invited to add your own R content feed to this site (Non-English R bloggers should add themselves- here)

Previous post presented the problem of dishonest casino that ocassionally uses loaded die. Sequence of the real states is hidden, and we are trying to figure it out just by looking at the observations (symbols).

# backtracking algorithm
for (i in 2:length(symbol.sequence)) {
# probability vector stores the current emission with respect to (i-1) observation of selected state and transition probability
# state vector (pointer) on the other hand is only storing the most probable state in (i-1), which we will later use for backtracking
tmp.path.probability <- lapply(states, function(l) {
max.k <- unlist(lapply(states, function(k) {
prob.history[i-1, k] + transition.matrix[k, l]
}))
return(c(states[which(max.k == max(max.k))], max(max.k) + emission.matrix[symbol.sequence[i], l]))
})
prob.history <- rbind(prob.history, data.frame(F = as.numeric(tmp.path.probability[[1]][2]), L = as.numeric(tmp.path.probability[[2]][2])))
state.history <- data.frame(F = c(as.character(state.history[,tmp.path.probability[[1]][1]]), "F"), L = c(as.character(state.history[,tmp.path.probability[[2]][1]]), "L"))
}
# selecting the most probable path
viterbi.path <- as.character(state.history[,c("F", "L")[which(max(prob.history[length(symbol.sequence), ]) == prob.history[length(symbol.sequence), ])]])

If we apply our implementation to the data in the previous post, we can get the idea how well can HMM reconstruct the real history.