We consider the problem of passive symbolic learning in the case of deterministic top-down tree transducers (DTOP). The passive learning problem deals with identifying a specific transducer in normal form from a finite set of behaviour examples. This problem is solved in word languages using the RPNI algorithm, that relies heavily on the Myhill-Nerode characterization of a minimal normal form on DFA. Its extensions to word transformations and tree languages follow the same pattern: first, a Myhill-Nerode theorem is identified, then the normal form it induces can be learnt from examples. To adapt this result in tree transducers, the Myhill-Nerode theorem requires that DTOP are considered with an inspection, i.e. an automaton that recognized the domain of the transformation. In its original form, the normalization (minimal earliest compatible normal form) and learning of DTOP is limited to deterministic top-down tree automata as inspections. In this talk, we show the challenges that an extension to regular inspections presents, and present two concurrent ways to deal with them:

1. first, by an extension of the Myhill-Nerode theorem on DTOP to the regular case, by defining a minimal *leftmost* earliest compatible normal form.
2. second, by reducing the problem to top-down domains, by using the regular inspection as a lookahead

The merits of these methods will be discussed for possible extensions of these methods to data trees.

tba

TBA