The Complexity of Learning Branches andStrategies from

We study the problems of eeciently learning innnite branches for nite state trees and winninig strategies for closed nite-state games using membership, and branch or strategy queries, respectively. We show that generally no eecient branch learning algorithm exists but we provide such algorithms for several natural cases, in particular for deadend free nite-state trees, the class of trees such that the set of innnite branches has positive measure, and several classes of mod-ulo trees. Furthermore, we nd a way to apply Angluin's results about the identiication of deterministic nite automata from queries, which yields positive and negative strategy learning results, in particular, we show that the class of deadend free closed nite-state games is eeciently strategy learnable from membership and strategy queries.