Learning from conflicts in real world environments for the realization of Cognitive Technical Systems

In this contribution, a novel learning method realizing the refinement of a Cognitive Technical System's pattern recognition and attention capabilities is presented. The method is implemented within a cognitive architecture with a representational level based on Situation-Operator-Modeling and high-level Petri Nets. Through the representational level, it is possible to realize a mental model mapping the complex structure of the real world internally in a compact format reduced to the relevant aspects. The mental model can be created and modified automatically by learning from interaction. If the perceived real world does not correspond to the system's mental model, the system detects ambiguities (or conflicts) inevitably. Then, the system tries to solve the conflicts by a more detailed view to the measured sensor inputs. Thus, new significant features (on a high abstraction level) can be derived from the measurements and taken into account to distinguish different (before apparently equal) situations. The contribution describes the proposed method and its fundamentals in detail. Furthermore, the realization of a cognitive mobile robot is presented as an application example illustrating the proposed method.