Brzozowski's algorithm for converting a DFA into an equivalent minimum-state DFA is remarkably simple: if $R(D)$ denotes the NFA formed by reversing all the edges in a DFA $D$, making the old start state an accepting state, and making the old accepting states start states, and if $P(N)$ denotes the result of applying the subset construction to the NFA $N$, then $$P(R(P(R(D))))$$ is a minimum-state DFA with the same language as $D$.
We can think of a DFA as a computational device that accepts an input string $w$ and then outputs 0 if $w$ ends in a rejecting state and 1 if $w$ ends in an accepting state. A natural generalization of DFAs associated each state in the DFA with some natural number between 0 and $k-1$, inclusive.
To the best of my knowledge, it's possible to minimize these modified classes of DFAs by using a distinguishability-based minimization algorithm, such as the canonical one by Hopcroft. However, I can't see how it would be possible to adapt Brzozowski's minimization algorithm to this new class of automata because the key step (reversing the automaton) no longer has a clear interpretation in this generalized setting.
Is there a known generalization of Brzozowski's algorithm for minimizing these sorts of automata? If not, are there any theoretical reasons why we'd expect that such a modified algorithm wouldn't exist?