The CONNECT Integrated Project aims at enabling continuous composition of networked systems, by developing techniques for synthesizing connectors. A prerequisite for synthesis is to learn about the interaction behavior of networked peers. The role of WP4 is to develop techniques for learning models of networked peers and middleware through exploratory interaction. During Y1 of CONNECT, exploratory work was performed to understand the requirements for learning techniques in the CONNECT process, and to develop concepts for using a priori knowledge about component interfaces as a basis for learning. During Y2, a major goal has been to develop techniques for automatically learning models of networked peers and middleware, based on the concepts developed during Y1 (cf. D4.1). This deliverable surveys the significant progress made on problems that are important for realizing this goal. We have developed techniques for drastically improving the efficiency of active learning, meaning to explore significantly larger parts of component behavior within a given time. We verified the power of our approaches by winning the ZULU challenge in competition with several very strong groups in the language learning community. We have also made significant breakthroughs for learning of rich models with data. We have developed a novel compact and intuitive automaton model for representing learned behavior in a canonical way. Canonicity is very helpful for organizing the resuls of exploratory interactions, since it allows the extension of stable techniques for learning classical finite-state automata (such as L ) to richer models with data. We confirm this by using the new automaton as a basis for an extension of L to richer automaton models. During Y1, we have identified abstractions as an important concept in the process of learning behavioral models of realistic systems: we introduce a method for refining a given abstraction when needed during the learning process. We also present a specialization of learning, which does not generate complete models of behavior, but concentrates on aspects that are relevant in a specific context. Finally, the deliverable reports on our efforts to learn non-functional properties, the development of a monitoring infrastructure, and further development of the learning tool infrastructure, based on LearnLib.