Learning Multiplicity Automata from Smallest Counterexamples

We show that multiplicity automata (MAs) with size n and input alphabet Σ can efficiently be learned from n(n+1)jΣj+2 smallest counterexamples. This improves on an earlier result of Bergadano and Varricchio. A unique representation for MAs is introduced. Our algorithm learns this representation. We also show that any learning algorithm for MAs needs at least 1/64n2|Σ| smallest counterexamples. Thus our upper bound on the number of counterexamples cannot be improved substantially.