A Comparison of Selected Variable Ordering Methods for NFA Induction

Improved SAT models for NFA learning

“…Many works on this topics are only formal and experimental results are also difficult to compare. For examples, the authors of [8,9] focus on a parallel solver for optimizing k. In [10], experiments are based on samples issued from the Waltz-DB database [2] of amino acid sequences, i.e., all the words are of size 6, and there cannot be any prefix word: in the tests we performed, only anagrams could be used in multisets. Moreover, for all the 50 instances we tried issued from this database, the M k model could be generated and solved in a reasonable time, without need of any model improvement.…”

“…. Then, if w ∈ S + , using several times Property 4, Constraints (4), (5), and ( 6) can be replaced by Constraints ( 8), (9), and (10) where l 0 = q 1 and N = [1, . .…”

“…. ∨ ¬d n ∨ ¬f j ) (9) i∈N,li∈K\{lj |1≤j<i} i∈N,di∈Du i ,q l i −1 ,q l aux w,l1,...,ln (10) Similarly, if w ∈ S − , using several times Property 4, Constraints ( 7) can be replaced by Constraints (11):…”

“…Parallel solvers for minimizing the inferred NFA size are presented in [8,9]. The author of [10] proposes two strategies, based on variable ordering, for solving the CSP formulation of the problem.…”

Optimized models and symmetry breaking for the NFA inference problem

“…Whereas some approaches focus on improving the solver, such as specific strategies [13], or parallel techniques [2] for optimizing k, we have an othogonal approach that consists in improving the model and using a standard solver. In [12], we studied complex data structures to generate smaller instances, and in [1], we proposed various models, and we use metaheuristics to optimize instance sizes.…”