Learning Several Languages from Labeled Strings: State Merging and Evolutionary Approaches

TL;DR

This paper introduces two novel methods for learning multiple disjoint DFAs from labeled strings, one using state merging heuristics and the other employing a multi-objective evolutionary algorithm, evaluated on industrial data.

Contribution

It presents new approaches combining state merging heuristics and evolutionary algorithms for learning multiple DFAs from labeled data.

Findings

State merging method produces effective sub-DFAs.

Evolutionary approach balances accuracy and number of DFAs.

Methods perform well on industrial datasets.

Abstract

The problem of learning pairwise disjoint deterministic finite automata (DFA) from positive examples has been recently addressed. In this paper, we address the problem of identifying a set of DFAs from labeled strings and come up with two methods. The first is based on state merging and a heuristic related to the size of each state merging iteration. State merging operations involving a large number of states are extracted, to provide sub-DFAs. The second method is based on a multi-objective evolutionary algorithm whose fitness function takes into account the accuracy of the DFA w.r.t. the learning sample, as well as the desired number of DFAs. We evaluate our methods on a dataset originated from industry.