Lots of work in Learning from Demonstration (LfD) have been done with the purpose of learning low-level motion primitives. In contrast to those approaches, here the possibilities of the high-level composition of primitives using dynamic behavior trees have been investigated. Starting with the set of python primitives, in order to mimic the behavior, optimal tree structure and optimal activation pattern laws need to be revealed, where the optimality is defined with respect to the standard LfD loss. Our focus here is not on the low-level learning of individual motions, rather on learning high-level action composition. We start from the precoded set of primitives and map a demonstration to a tree of this primitives. The aim is not to reproduce every demonstration one can imagine, rather demonstration from the finite set of given primitives. The results have shown, that the learned tree is capable to reproduce, besides the behavior which has been demonstrated, also the high-level behaviors which haven’t been demonstrated. Therefore, the learned behavior tree is capable to successfully adapt to the unseen situations, which accounts to the inherent reactive feature of the dynamic behavior trees.