Approximate learning of random subsequential transducers