PA learning

[pberko]

Hello

Do you have a working example for pa_learning?

Thanks

[vhavlena]

Hello,
an example of the input file you can find on https://github.com/matousp/datasets/tree/master/scada-iec104/iec104-traffic

[pberko]

Thanks

The file looks different from what was in my mind.
I thought it should contain paths from a DFA alphabet.
how is this file extracted from DFA?

[vhavlena]

It is quite tailored for our application. The input file is basically a long sequence of symbols. In the first step based on an expert knowledge, these symbols are put together into strings and in the second step such a multiset of strings is then the input for DPA learning.

[pberko]

So in line 66 I get list of strings which is actually automata paths?

[pberko]

@vhavlena
Another question
the function
""" Add string to frequency prefix tree """ def add_string(self, string, label=0): act = self._root self._ini[act] = self._ini[act] + 1 for i in range(len(string)): try: self.flanguages[act][tuple(string[i:])] += 1 ....

is recieving only tuples as input (as in your example) or can also get simple string i.e. "a a a"

[vhavlena]

1. Yes. The list of strings is the input for learning. In your output, each line (list) is a string where single pair of strings is a symbol.
2. The variable string should be a list of arbitrary symbols; you can give there arbitrary string (in your case "aaa").

[pberko]

Hello @vhavlena

I tried the algorithm with other samples but I'm afraid I did a mistake since the results are different from expected:

The apleh-bet is "a, b"
input file with paths generated from a "blackbox automaton".
input file :https://github.com/pberko/detano/blob/master/train1clean.csv
https://github.com/pberko/detano/blob/master/blackbox111.pdf

I use the file as input file to pa_learning but I got the output:
https://github.com/pberko/detano/blob/master/graphviz%20(2).pdf
which looks different

[vhavlena]

Hello @pberko,

Not sure if you are doing something wrong. I think there are several issues to consider. Your black-box automaton is Markov chain, right? In that case you likely get something different, because DPA on the output has accepting probabilities that may cause some bias for your setting (maybe you can try specialised algorithms for learning MCs). The second issue are learning parameters: different parameters may lead to different automata.