Traditional algorithms of machine learning implemented in cognitive architectures demonstrate lack of autonomous exploration in an unknown environment. However, the latter is one of the most distinctive attributes of cognitive behaviour. The paper proposes an approach of self-reinforcement cognitive learning combining unsupervised goal acquisition, active Markov-based goal attaining and spatial-semantic hierarchical representation within an open-ended system architecture. The novelty of the method consists in division of goals into the classes of parameter goal, invariant goal and context goal. The system exhibits incremental learning in such a manner as to allow effective transferable representation of high-level concepts.