The maximum likelihood estimate

Another empirical probability measure that can be used is the maximum likelihood estimate:

$$P(s)=\frac{n_s}{N_s}$$

This estimate is frequently used because of its simplicity, and because it generally gives correct results.

It was used by Guyon and Pereira [39] to train a variable length Markov model on strings. It has also been used by Galata, Johnson and Hogg [33] to train a variable length Markov model on the pathlet models derived from flow vectors.