Learning deterministic even linear languages from positive examples

We consider the problem of learning deterministic even linear languages from positive examples. We show that, for any nonnegative integer k, the class of LR(k) even linear languages is not learnable from positive examples while there is a subclass called LRS(k), which is a natural subclass of LR(k) in the strong sense, learnable from positive examples. Our learning algorithm identifies this subclass in the limit with almost linear time in updating conjectures. As a corollary, in terms of even linear grammars, we have a learning algorithm for k-reversible languages that is more efficient than the one proposed by Angluin.